Julia: What MIT Actually Gave Away, and What JuliaHub W...

I gave a small machine shop a week of work to plan. The jobs are invented, generated by a deterministic sequence written into the script so anyone running it gets the same numbers, and everything below follows from them. Sixty open quotes, four work centers, a fixed number of hours at each, four repeat customers whose jobs get run regardless of margin. Which quotes do you accept?

The answer came back in 0.276 seconds, proved optimal, gap zero. Thirty-one of the sixty jobs, 78,240 dollars and 44 cents of margin, every work center between 95.7 and 99.8 percent booked. Taking the fattest-margin jobs first, which is what most shops actually do, gets you 70,711 dollars and 80 cents. The difference is 7,528 dollars and 64 cents on one week, or 10.65 percent, and it exists because capacity is shared across centers and greed at the top of the list starves the bottom.

The whole thing cost nothing. Julia is free. The modeling layer is free. The solver is free. The total software bill for that answer was zero dollars, and the total permission required was zero.

Then I tried to find out what the commercial version of this world costs, and that is where the article is.

On 31 August 2026, MIT News published "How an MIT research project became a global programming language", a warm piece about Julia, the language four people started complaining their way into around 2009. It is a good story and most of it is true. It also contains nine links in its body, eight of which go to juliahub.com, and it does not mention a price anywhere. Neither does juliahub.com, for the product the article is mostly about.

I have written before about AFSIM, the Air Force simulation framework Boeing developed and funded through its own internal research and development between 2003 and 2014, and delivered to the government under contract in February 2013 with unlimited rights including source code. AFSIM is described by the government's own information analysis center as "government-owned, open-sourced, and community-informed." Nobody can download it. There is a documented way in, which is an Information Transfer Agreement signed with AFRL's Aerospace Systems Directorate, and RAND, writing for the Air Force, put the licensee base at more than 275 government, industry and academic organizations, giving no source for the count. What there is not, and cannot be, is a link.

Julia is the mirror image. No export license attaches to the software itself, sanctions law being a separate question, so in practice anybody can download all of it right now, and the finished product built on top has one button where a price should be. Both stories are about the same question, which is the only question a small business ever really asks about software: how far do I get without asking permission?

For Julia the answer turns out to be surprisingly far, and the interesting part is exactly where it stops.

What MIT gave away, and what the license actually says

On 14 February 2012 four people published "Why We Created Julia" and it reads like a complaint that got out of hand. "We want a language that's open source, with a liberal license. We want the speed of C with the dynamism of Ruby." They admitted it was not finished: "It already delivers on 90% of our ungracious demands."

There is a footnote on that page now, added later, and it is my favorite thing about the Julia project. "This was originally written as being 'time for a 1.0 release' but it turns out this was a bit too greedy: Julia 1.0 wasn't released until considerably later." Considerably later was 8 August 2018, six and a half years on, and the headline feature was not speed. It was a promise not to break your code: "code you write for Julia 1.0 will continue to work in Julia 1.1, 1.2, etc."

The license is the load-bearing fact. LICENSE.md in the Julia repository is a plain MIT license, copyright 2009 to 2025, naming Jeff Bezanson, Stefan Karpinski, Viral B. Shah and other contributors. Permission is granted "to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software..." There is no field-of-use restriction. A business can use it, embed it in a product, and sell that product, and owes nobody a cent or a phone call.

One caveat that a lawyer will care about and most articles skip. The MIT license covers the language: the compiler, most of the standard library, the utilities. The distribution you actually download bundles other people's libraries, and the project's own THIRDPARTY.md lists them: GMP under LGPL3+ or GPL2+, MPFR under LGPL3+, several SuiteSparse modules at GPL-2.0+, libgit2 at GPL2+ with a linking exception. If you are shipping a closed binary to a customer, that list is the thing to read, not the headline.

Governance is where the story gets more interesting than the press coverage suggests. Julia's governance page describes roughly forty committers working by informal consensus with a standing triage call, a seven-person Stewards group that exists only to handle conduct complaints, and a fiscal sponsorship arrangement with NumFOCUS under which five named people, Tim Holy, Steven Johnson, Viral Shah, Stefan Karpinski and Avik Sengupta, have signed an agreement that "they, or their successors, will manage the technical direction and programs of the project within the scope of the organization's nonprofit status."

The same page carries a section titled "Related Organizations With No Governance Role," and the two organizations it names are the MIT Julia Lab and JuliaHub. That disclaimer is doing real work, and it is fair as far as it goes. Stefan Karpinski, writing for the company, has put the counterweight in print himself: JuliaHub "employs many of the top contributors to Julia itself (half of the top ten contributors by commits)." Today one person is simultaneously JuliaHub's VP of Modeling and Simulation, Pumas-AI's Director of Scientific Research, and, by JuliaHub's own description, a Co-PI of the MIT Julia Lab. The separation is legal and structural. It is not a separation of people.

The other kind of open source

Here is the sentence that explains why AFSIM and Julia can both be called open source and mean opposite things.

When the government holds unlimited rights in software, DFARS 252.227-7014 says exactly what that means: "rights to use, modify, reproduce, release, perform, display, or disclose computer software... in any manner and for any purpose whatsoever, and to have or authorize others to do so." So as a matter of data rights, the Air Force could hand AFSIM's source to anyone it likes, including a competitor of the company that wrote it.

It does not, because data rights and distribution rights are two different locks. DSIAC's AFSIM page opens by calling the framework "government-owned, open-sourced, and community-informed" and then, several screens down, states that "the core AFSIM framework source code, binaries, and extensions cannot be redistributed or shared without prior coordination and approval," marked Distribution Statement F. The current authorised wording of that statement, from DoD Instruction 5230.24 dated 10 January 2023, is "Further distribution only as directed by [controlling DoD office] [date of determination] or higher DoD authority." The same page adds a line worth sitting with: "All output from AFSIM is also export controlled, regardless of input."

Now the resolving fact, which took a while to find and is the cleanest version of this whole article's argument.

Under the Export Administration Regulations, published software is not restricted software. It is not software subject to the rules at all. 15 CFR 734.3(b)(3) puts information and software that "are published, as described in § 734.7" outside the scope entirely, and 734.7(a) lists "posting on the Internet on sites available to the public" as one of the ways to get there. Nobody signs off. Publishing is self-executing: you publish, and the control simply stops attaching.

The other regime works the opposite way. Under the ITAR, 22 CFR 120.34 lists eight routes into the public domain, and the one for public release, at (a)(7), is open only "after approval by the cognizant U.S. Government department or agency." AFSIM's own marking names both regimes in one breath, as boilerplate notices do, and states no munitions-list category and no export classification number, so which regime governs it is not something that page decides.

That is the asymmetry, and it is not about which codebase is more dangerous. Under one regime a private party publishes and the restriction ceases. Under the other, no amount of publishing does anything at all until a government office signs. Julia is not free of export law because MIT wrote a permissive license. What matters is that it was published, by people who were allowed to publish it, and that nobody had to approve that. One honest caveat on the reading: 734.7 conditions the exclusion on material being available "without restrictions upon its further dissemination," and an open-source license does impose one condition, which is keeping the copyright notice. The practical reading in the field is that this is not the kind of restriction the rule is aimed at. I did not find a regulator's statement saying so, and I am not going to invent one.

The 2024 SHARE IT Act does not change this, and it is worth saying so because the name invites the opposite reading. Public Law 118-187, approved 23 December 2024, requires agencies to store custom-developed code in "not less than 1 public repository or private repository" accessible to federal employees. A private repo satisfies it. It is a code-sharing-within-government law, not an open-source mandate. Its export carve-out is not automatic either: the Act files a flat exemption for classified code and national security systems under "Automatic," and puts the export-control one under "Discretionary," where an agency's chief information officer, in consultation with the Federal Privacy Council, may exempt code the agency is barred from sharing. And it applies only to code developed or revised at least 180 days after enactment, which is to say from about June 2025 onward.

One update to my own earlier reporting. When I wrote about AFSIM in July 2026, I noted that the trade-reported 85 million dollar ceiling for the AFSIM FORGE contract vehicle was unconfirmed, because the federal spending data showed no base-and-all-options value. It shows one now. As of 6 September 2026 the USAspending API returns a base-and-all-options figure of 85,000,000 on each of the three awards, to Radiance Technologies, Stellar Science and Infinity Labs, all running to 20 August 2029. Whether that is one shared ceiling or three separate ones the data does not say, and total obligations on all three are still zero, which for an indefinite-delivery vehicle means the money moves on task orders that have not posted yet.

The MIT News piece, checked link by link

The MIT News article has nine links in its body. Eight go to juliahub.com. One goes to YouTube. Not one independent source is linked in the body, though the page's Related Links block does point at several MIT domains. That is not a scandal, it is a house style for institutional news writing, but it does mean every claim in it arrives through one company's material.

Two of the claims do not survive a check.

The aircraft collision one. MIT News says researchers "used Julia to create a program for avoiding aircraft collisions. They found it ran about 50 times faster than an earlier version built on Python." The article links that sentence to a JuliaHub case study. The words "Python" and "faster" do not appear on that page. No speed multiple of any kind appears on that page. What it says is "Julia reduced the time required to conduct these computations by several years," and separately that "Previously, FAA partners such as Lincoln Labs needed to use Matlab to develop their algorithms and then program in C++ in order to run the algorithms over very large datasets."

The primary source underneath is a 54-slide JuliaCon 2015 deck by Robert Moss, titled "Using julia as a Specification Language for the Next Generation Airborne Collision Avoidance System," work its own title slide says was sponsored by the Federal Aviation Administration. It mentions MATLAB three times and Python zero times, and its performance line is "Julia computes exhaustive comparison over 6.5e11 discrete decision points. Would be computationally expensive for MATLAB." The comparison in the primary record is against MATLAB, not Python, and it is not a multiple. I could not find a primary source for the 50x anywhere. I am not saying anyone invented it. I am saying the page MIT News chose to link contains neither the number nor the language it attributes, and the sentence carries no other citation.

The WhatsApp one. MIT News says Meta engineers used Julia to build "a better audio codec for WhatsApp's 4 billion users." Meta's last disclosed WhatsApp figure was more than 3 billion monthly users, stated by Mark Zuckerberg on the company's first-quarter 2025 earnings call in May 2025. No source at any tier reports 4 billion. That sentence carries no link at all.

Two claims that do survive, with caveats worth having.

Moderna. MIT News says a pharmaceutical modeling platform built in Julia "was used to accelerate development of Moderna's Covid-19 vaccine." The platform is Pumas, the work is peer-reviewed, and it is real. The disclosure attached to that paper matters: two Pumas-AI authors "were paid by Moderna, Inc., to perform the analysis," on a study funded by Moderna and BARDA with a Moderna corresponding author. That is normal contract pharmacometrics. It is not independent validation of a tool, and it should not be read as one.

The black holes. Astronomers do use Julia for black-hole imaging, through Comrade.jl, and its peer-reviewed JOSS paper says the package "has already been widely used within the Event Horizon Telescope Collaboration." The widely-quoted "100x faster than other EHT modeling packages" is not in that paper. It is a line in a ten-minute lightning-talk abstract, with no named baseline and no published benchmark. It is better attributed than the aircraft claim, because a researcher put his own name to it. It is not better evidenced.

And the number everyone repeats. JuliaHub's homepage stat block reads "1M+" over the word "Users," and MIT News renders that as "more than 1 million users." It appears only on JuliaHub marketing surfaces, never on julialang.org, with no methodology, no definition of a user, and no measurement date on any page I opened. The registry count is checkable and I checked it: the General registry's Registry.toml listed 14,284 packages on 6 September 2026, which is more than either the "over 12,000" on julialang.org or the "13K+" on juliahub.com. Both of those are floor claims and both are true. They are also conservative and stale, and the registry is append-only, so 14,284 is "ever registered and not removed," not "alive." So I counted the living ones. Cloning the registry and asking which packages registered any version at all in the twelve months to 6 September 2026 gives 5,652, which is 39.6 percent. Stripping out the 1,825 auto-generated binary wrappers barely moves the ratio, at 4,877 of 12,459 hand-written packages, or 39.1 percent. Tighten the window and it falls fast: 4,119 packages in six months, 2,963 in three. And that is an upper bound, because a single patch bump counts as alive. The honest figure for a business asking what it can depend on is closer to five thousand than fourteen.

Three of the six numbers on JuliaHub's homepage are not in the pages they cite

The juliahub.com homepage carries six result tiles. I opened the page each one links to.

Three of the six check out at the headline number. Instron's 500x appears five times in the linked case study. Mitsubishi Electric Research Laboratories' "<2% error" appears verbatim. NASA's 15,000x is on tape, and the arithmetic works: 900 seconds to 0.060 seconds is exactly 15,000.

Three of the six do not appear in the page they link to, or in the talk behind it.

ASML. The tile says "50% reduction in development cycle time across 700+ engineers." In the keynote it points at, an ASML vice president says "we're embarking in Julia. I should it's out of the what was it 80 million lines of code I don't think we have a million yet," and a colleague adds "we're piloting we're figuring out." Both are quoted from the machine transcript, disfluencies and all. A pilot in a codebase where, in his own words, they do not have a million lines of Julia out of eighty million is a different fact from a 50 percent cut in development cycle time across 700 engineers.

Boeing. The tile reads "Boeing built a guidance, navigation, and control system in 2 years instead of 5 years," attributed to "Boeing Keynote JuliaCon 2024." The conference schedule shows a 25-minute session titled "JuliaSim: Bringing Julia to Industrial Modeling and Simulation," with three speakers: Chris Rackauckas, Viral B. Shah and Gary Mansouri. Two of those are JuliaHub principals. The abstract describes the Boeing participant, Gary Mansouri, as joining "for a conversation." The two-versus-five-years figure is not on the landing page and not in the talk.

Williams Racing. The tile says "50x more accurate prediction for its F1 car velocity and angle, with 4x faster execution." Neither number appears in the case study. Williams' own hosted copy of the same case study carries a different stat block entirely: 169x faster, 7 percent more accurate, 1000x speed improvement, 2.3x higher fidelity. No 50x. No 4x. Williams' copy also names JuliaSim eight times and Dyad zero times, and mentions MATLAB, the named baseline, zero times, where JuliaHub's copy mentions MATLAB five times.

Two more things I found while checking, both of which are the more interesting finding.

The NASA number is the cleanest of the six, and the engineer who produced it plays it down on camera. Jonnie Diegelman, a guidance navigation and control engineer at NASA's Launch Services Program, in a talk JuliaHub dates to March 2021: "that 15 000 number doesn't matter as much as just the fact that it's very very fast now and we're still incorporating some features maybe maybe it'll take twice as much time when we're done." He names his baseline, which is his team's own Simulink tool, and states the workload, which is 514 Monte Carlo runs. That is exactly how a speed claim should be made, and the marketing tile that repeats it drops the baseline, drops the workload, and describes the work as "mission planning" when the talk is about spacecraft stage separation dynamics.

And the Instron case study contains a units error that runs in the vendor's own disfavor. The document says a loss-function evaluation takes "10+ minutes" and the optimization needs "1000+ iterations." That multiplies to 166.67 hours. The case study prints it as "well over 166 days." The homepage's "days to minutes" is what you get from the correct figure divided by 500. The case study's "months to hours" is what you get from the erroneous one. This is not inflation. It is a slipped unit that has been sitting in a customer-facing document for years.

One last thing about that case-study index. Fifteen case studies carry five distinct publication dates between them, thirteen of them share three dates, and six are stamped 2 January 2023. Those are content-migration timestamps, not publication dates, which means the freshness signal on that whole library is worthless. The Instron study is stamped 2 January 2024 and describes a webinar from November 2023 and a talk from July 2022.

The benchmark in the funding announcement, against the paper it links to

This is the one worth reading slowly, because it is checkable in two documents and it goes further than a rounded number.

JuliaHub's Series B announcement of 30 April 2026 makes a specific claim about competing AI systems:

"In recent agentic benchmarking for chemical process modeling, general LLM systems such as Codex, Claude Code (Opus), and Gemini barely completed the initial setup. Dyad almost entirely automated the whole process of creating model-predictive controllers to optimize yields of a chemical plant."

The phrase "recent agentic benchmarking" is a hyperlink, and it points at a real peer-reviewed paper: "Improving process systems engineering with specialized multi-agent large language models," in Chemical Engineering Journal Advances, volume 26, May 2026. Here is that paper's own abstract, as the journal indexes it:

"Two multi-agent systems (Dyad 2.1 and Claude Opus 4.6) were compared with two single-agent baselines (ChatGPT 5.2 and Google Gemini Pro 3)... All LLMs converged to latent-variable linear chemometric models and achieved near-linear parity with R2 close to unity across training, validation, and test sets... all systems reconstructed a moment-based population balance model... All NMPCs were implementable at a 1 min sampling time, achieved near set-point tracking with moderate control effort over five scenarios, and Dyad provided the best overall closed-loop performance."

Three things come apart there.

Every system finished. The abstract says all of them converged in the first case study, all of them reconstructed the model in the second, and all of the controllers were implementable in the third. "Barely completed the initial setup" and "all NMPCs were implementable" cannot both describe the same experiment. There is one sub-case that goes the vendor's way, and it belongs here: on crystallization, the abstract says only the multi-agent models reproduced the dynamics consistently and were kept for the next stage, so the two single-agent baselines did fall short there. One sub-case is not the initial setup.

Claude was not a baseline. The paper puts Claude Opus 4.6 in the same category as Dyad, as one of the two multi-agent systems, and the single-agent baselines it is measured against are ChatGPT and Gemini. The announcement lists it among the general systems that failed.

Codex is not among the four systems the abstract names. Those are Dyad 2.1, Claude Opus 4.6, ChatGPT 5.2 and Gemini Pro 3. The announcement names a product that does not appear in the evidence it cites.

What the paper actually concludes about Dyad is real, and narrower: among four controllers that all worked, Dyad gave the best overall closed-loop performance, and the multi-agent workflows held up better under dataset shift. That is a genuine result, and it is a long way from the sentence in the funding announcement.

Some fairness, because this deserves it. Eight of the paper's thirteen authors carry a Julia Computing affiliation, which is disclosed on the paper. The funding registered against it is Brazilian public money, from CAPES, FAPERJ and CNPq, and JuliaHub does not appear as a funder, so this is not a study a vendor bought. I could not read the paper's competing-interests declaration, so I am not saying anything about whether or how the affiliations were declared beyond the author list, which states them plainly. And there is a charitable explanation available for how the announcement got where it did: the paper's own graphical abstract puts three machine-generated marks loosely evoking commercial chat assistants under a "General-Purpose LLMs" heading, opposite Dyad, while the paper's text calls Claude a multi-agent system. The marketing tracks the picture rather than the sentence. That figure is itself machine-made, down to a misspelled axis label, so I am not going to say which company each mark is meant to be.

Two more things about that announcement. It says Dyad compresses R&D "from months to mere days" in its opening line and "from months to minutes" a few paragraphs later, neither figure attached to a customer or a measurement. And the benchmark is of Dyad 2.1 while the release announces Dyad 3.0, with every system frozen at its early March 2026 state, so "Claude Opus 4.6" there is a specific dated system tested six months ago rather than anything current.

One small irony, since this article is about who is allowed to read things. That paper is gold open access under a Creative Commons license, which is to say free for anyone to read and redistribute. I could not open its body text through any ordinary route, and everything above comes from the journal's own indexed abstract, the publisher's open image server, and the metadata registries. The license says open. The delivery does not.

What it costs, which is the part that changed

Here is the sequence, from the archives, and it is the most useful thing in this article for anyone evaluating JuliaHub.

In September 2023 JuliaHub published a full three-tier price list under a heading that read, in its own words, "Complete Transparency, Full Control." Standard came with "20 hours of free compute that resets every single month," usage-based after that. Team started at 10,000 dollars. Enterprise started at 35,000 dollars. Each additional user was "250 per Month or 2K per Year." Cluster compute was priced per thread hour.

By the last archived priced snapshot, 22 June 2025, Team had been repriced from "Starting at $10,000" to "$25,000 +" for up to five users, and Enterprise's dollar figure was gone, replaced with "Contact Us."

Sometime between 22 June 2025 and 6 October 2025, the prices went away entirely. The old URL now redirects to a page headed "Request JuliaHub Pricing" whose entire content is a contact form: first name, last name, email, organization, position, country. There is not a dollar figure on it. The technical documentation had its own pricing page, which quantified the free tier precisely at 20 free hours a month on a 2 vCPU machine and 10 cents per thread hour after that. That page is gone too, and "Pricing" no longer appears in the documentation navigation.

The free tier still exists. The registration page at auth.juliahub.com asks for an email address and nothing else, with no credit card field and no sales gate, though I did not complete a signup to see what lands on the other side. What it entitles you to is no longer written down anywhere. The word on the live site is "free," with no number attached to it.

Now the part that matters most, because it is the part a small business will get wrong.

Dyad, JuliaHub's flagship, does publish prices. Four of them: Dyad Free at 0 dollars a month, Dyad plus Light AI at 20, plus Medium AI at 60, plus Heavy AI at 200. That looks like ordinary self-serve software, and the buttons do lead to a real in-product payment route.

The heading above that price table reads "Non-Commercial Use. For researchers, students, and engineers building on their own." The EULA never names Dyad, so here is the chain rather than an assertion. Dyad Studio's own listing points its license at juliahub.com/company/eula, and that URL redirects to the JuliaHub end user license agreement, last modified 3 August 2026. That document makes it binding:

"Permitted Purpose" means: (a) during the first thirty (30) days following Customer's initial access to the Licensed Software (the "Trial Period"), Customer's internal business purposes, including commercial use for evaluation purposes; and (b) following expiration of the Trial Period, non-commercial and academic purposes only.

The next sentence is in capitals in the original, and says that after the trial period the customer may not use the software for any commercial purpose whatsoever without first entering into a separate order form or license agreement.

So if you run a business, download Dyad Free, and use it on your own product work, you are inside your license for thirty days and outside it on day thirty-one. The price of the thing you would then need is not published. It is a second contact form on the same page, under the heading "Commercial Use."

Two smaller details from the same document set, both of which tell you something. The license bars the customer from using the software, any output, or derived data "to develop, train, fine-tune, validate, or improve any artificial intelligence, machine learning, deep learning, or other algorithmic, statistical, or inference-based model or system," while the customer-data clause grants JuliaHub a royalty-free license to use customer data. Neither of those is unusual; restrictions on benchmarking and on training models from output, and broad grants over customer data, are close to standard in commercial software terms, and it would be unfair to present them as something this company invented. And the separate Dyad Source Available License, which governs some of the JuliaHub-authored Julia packages, treats government use as a commercial purpose and does not permit it. That license is derived from PolyForm Noncommercial 1.0.0, which expressly permits government use: "Use by any charitable organization, educational institution, public research organization, public safety or health organization, environmental protection organization, or government institution is use for a permitted purpose regardless of the source of funding." JuliaHub did not add a restriction to a silent license. It inverted an explicit permission, and dropped the charities and public research organizations along with the government.

Is it fair to single out quote-only pricing? Mostly no

I went looking for the counter-argument and found it immediately.

Ansys publishes no price. Its old pricing URL now redirects to a page that no longer exists, and its most self-serve-looking product, Discovery, carries no dollar figure. Dassault Systèmes publishes no price for Dymola. COMSOL publishes no price for anything and frames the quote gate as a feature. Gurobi's pricing page contains zero dollar figures and one call to action, which is "Request a Quote." Siemens is the partial exception, and an earlier version of this paragraph had it wrong: its Simcenter Amesim product page names no price, but Siemens' own storefront sells a named Amesim configuration for 4,272 dollars with an add-to-cart button. So the category is a spectrum rather than a wall, and most of it sits at the far end.

So JuliaHub is not the exception. It is behaving exactly like the category. What makes its case sting is that it used to publish prices under a heading promising transparency, and stopped.

The genuine outlier runs the other way, and it is the forty-year-old proprietary incumbent. MathWorks publishes a per-product price on a public page with a cart button, and a perpetual price list as a downloadable PDF dated March 2026, eleven pages of products in two license-count columns. Not quite every row carries a number: about ten are blank, and two of those are the Copilot products, which say to visit the sales page. So even the transparent vendor has quote-only lines. It publishes the list anyway, which is more than the rest of them do. MATLAB is 1,050 dollars a year on an annual individual license, or 2,625 perpetual. Simulink is 1,585 a year, or 3,965 perpetual. Simscape, which is the acausal physical-modeling product that competes most directly with Dyad, is 1,050 a year, or 2,625 perpetual, and it requires both MATLAB and Simulink as prerequisites, so the honest entry price for that stack is 3,685 dollars a year for one commercial seat, not 1,050.

MathWorks also runs an eligibility gate, just a different kind. The Startup Suite is 4,235 dollars a year and the button says Apply now rather than Add to cart, because it is "for use at approved early-stage companies." And the cheapest MATLAB and Simulink bundle in existence is the Home Suite at 165 dollars, with Simscape and sixty-odd other add-ons sold separately on top of it, whose license forbids exactly the person reading this: not for "government, academic, commercial, for-profit, revenue-generating, or other organizational use."

That is worth sitting with. A small manufacturer is locked out of the cheap tier by license text rather than by price. It is the same gate as a contact form, enforced with a sentence instead of a sales call.

What a small business can actually run this week

This is the part nobody writes, so here are numbers I measured on a machine rather than numbers a vendor gave me.

The Julia 1.12.7 Windows build downloaded as a 280 MB archive and expanded to 1.1 GB on disk. Installing the two packages I needed took four and a half minutes wall clock, of which 136 seconds was precompilation of 39 dependencies. That is the entire setup, and it is the whole bill.

The shop-week model is 40 lines. It reads like the problem:

using JuMP, HiGHS

model = Model(HiGHS.Optimizer)
@variable(model, take[1:NJOBS], Bin)

# capacity: hours booked at each center cannot exceed the week
for c in CENTERS
    @constraint(model,
        sum(get(jobs[j].route, c, 0.0) * take[j] for j in 1:NJOBS) <= CAPACITY[c])
end

# the repeat customers get their work run, margin or not
for j in 1:NJOBS
    jobs[j].repeat_customer && @constraint(model, take[j] == 1)
end

@objective(model, Max, sum(margin(jobs[j]) * take[j] for j in 1:NJOBS))
optimize!(model)

JuMP is MPL-2.0. HiGHS is MIT. Both are free for commercial use with no seat count, no renewal and no trial period.

Then I asked how big the quote pile can get before the free solver struggles, holding the squeeze constant so capacity is always half of total demand:

Candidate jobs	Binary variables	Solve time	Result
60	60	0.055 s	optimal, gap 0
250	250	2.178 s	optimal, gap 0.01%
1,000	1,000	5.331 s	optimal, gap 0.01%
5,000	5,000	0.970 s	optimal, gap 0.007%
25,000	25,000	13.477 s	optimal, gap 0.002%
100,000	100,000	133.084 s	optimal, gap 0.002%

A hundred thousand binary decisions, settled in two minutes and thirteen seconds, on a free solver, on one machine. The non-monotonic row is real and I left it in: 5,000 solved faster than 1,000, because branch-and-bound is like that and anyone who tells you otherwise is selling something.

Now the honesty, because that table flatters the free layer. This problem class is easy. It is a knapsack with four constraints, and the structure is exactly what a modern solver eats. On genuinely hard problems the gap is real and measurable. On Hans Mittelmann's benchmark of hard mixed-integer instances, dated 7 July 2026, HiGHS solved 158 of 240 where the commercial COPT solved 219. On the linear programming benchmark, COPT solved all 65 instances and HiGHS solved 54. If your problem has routing with time windows, or sequencing, or big-M formulations, you will find the wall.

Two notes on that comparison, both of which cut against the commercial side. Unsolved instances are charged the full time limit, so the headline speed ratios overstate the gap. And the reason Gurobi, CPLEX and Xpress are absent from those tables is not that they lost. Mittelmann's own note says "IBM and FICO demanded that results for their solvers be removed" and "In August 2024 Gurobi decided to withdraw from the benchmarks as well." The commercial solvers are almost certainly still the fastest. You just cannot check it in the suite everyone cites, though independent head-to-head comparisons do still get published elsewhere.

The sharpest statement of where free stops is not in a benchmark table. It is one sentence in JuMP's own documentation, on the Traveling Salesperson tutorial, which is the page a delivery business would open to work out a driver's route: "We use Gurobi for this model because HiGHS does not support lazy constraints." The free stack's own route-planning example reaches for a commercial solver, and says why, in a line. To be fair to it, that same page solves the problem twice, and the first method, an iterative one that adds subtour constraints in a loop, runs on the free solver perfectly well; it is the more sophisticated callback approach that needs the paid one. The documentation does not pretend the gap is not there, and neither should anyone recommending it.

Deployment is the other question an engineering reader will ask, and the answer is better than Julia's reputation suggests. PackageCompiler.jl, which is MIT licensed and whose copyright line also names Julia Computing, will build an app bundle that runs on a machine that, in the documentation's own words, might not even have Julia installed, because the bundle brings Julia along. You do not have to ship your source files. That is not the same as protecting them: PackageCompiler's own documentation has a section on reverse-engineering the result, covering build paths recoverable from the binary and a bundled Julia that will happily load the image and reflect over it. What the documentation does not publish anywhere is how big the result is, and I am not going to invent a figure; the only absolute number in the primary material is a 1.1 MB trimmed hello-world in the Julia 1.12 release post. That trimming, the --trim flag, is explicitly experimental, and in its safe mode every dynamic call site becomes a compile-time error. Dynamic dispatch is idiomatic Julia and pervasive in exactly the packages a small business would use, so treat small standalone binaries as a research direction and app bundles as the thing that works today.

For physical modeling, the acausal component-based work that Dyad and Simscape both sell, the free layer is ModelingToolkit.jl and ModelingToolkitStandardLibrary.jl, and the license on that standard library is MIT, copyright Julia Computing. The same company that sells Dyad's component libraries publishes a free set under a permissive license with no trial clock. That is a real and generous thing and it deserves to be said plainly in an article that has been rough on the marketing.

So I built something with it. A shop wants to know how long it has after the heater quits on a winter night. In ModelingToolkit you do not write the differential equation. You declare the parts and say what is connected to what:

@named shop_mass  = HeatCapacitor(C = 2.0e7)      # air plus a 150 mm concrete slab
@named envelope   = ThermalConductor(G = 320.0)   # 690 m^2 of envelope plus infiltration
@named outside    = FixedTemperature(T = 268.15)  # a still minus 5 C night
@named heater     = PrescribedHeatFlow()
@named heater_cmd = Constant(k = 0.0)             # the failure: heater delivers nothing

eqs = [connect(heater_cmd.output, heater.Q_flow)
       connect(heater.port, shop_mass.port)
       connect(shop_mass.port, envelope.port_a)
       connect(envelope.port_b, outside.port)]

Five components and four connections went in. mtkcompile reduced the system to one unknown and one equation, which is the symbolic index reduction and simplification that Simscape and Dyad both sell as a headline feature. It cost nothing to run, and here this article's own trap closes on me: free to run is not the same as permissively licensed. ModelingToolkit's own license file is MIT and its base package is MIT, but since version 11 the index-reduction and tearing algorithms live in separate packages, StateSelection and ModelingToolkitTearing, which are AGPL-3.0 and are hard dependencies of the thing I installed. Anyone planning to embed this in a product they ship or host needs to read those two licenses and not the headline one. The answer, for those parameters, is a time constant of 17.4 hours and a drop through 15 degrees at 3.87 hours after the heater stops. The simulation and the closed-form check agree.

The honest cost of that: 38 seconds of wall clock, of which 3.2 seconds was the actual solve. The rest is Julia starting and the model compiling. On a laptop, once per session, and then it is instant. That gap is the single most-complained-about thing about Julia and this is what it looks like on a small model.

The other free option in that space is older and better known than most American small businesses realize. OpenModelica has been maintained since a consortium formed around it in December 2007, its current release is 1.27.0 from July 2026, and commercial use is expressly permitted by the consortium's own bylaws. It reads the Modelica Standard Library, which is BSD-3 licensed and contains 1,430 component models and blocks, 534 example models and 1,219 functions in version 4.1.0. So do Dymola, Simcenter Amesim, Modelon Impact, MapleSim and Wolfram SystemModeler. You can build a model in the free tool and hand the same model to a paid one later, which is the strongest portability argument available anywhere in this category.

One trap there, and it is exactly the kind this article exists to catch. OpenModelica's own summary page still describes its license as GPL version 3. The license text that actually ships, OSMC-PL version 1.8, makes the non-member mode GNU Affero GPL version 3, and the consortium's bylaws revision history records the swap. For a shop that only runs models internally, this changes nothing. For a shop that wants to embed OpenModelica in a product it ships or hosts, AGPL is materially different from GPL, and the way out is paid consortium membership. "Free and commercially usable" is true. "Free and commercially redistributable inside a proprietary product" is not, without joining.

What the free layer costs you instead

It costs you people, and the numbers are not flattering.

Julia sits at number 21 on the TIOBE index for September 2026 with 0.74 percent, against Python at number 1 with 17.76 percent. TIOBE counts search results, not employment, and its editor's headline that month is bullish, asking whether Julia is going to re-enter the top 20 and noting that Julia "is one of the languages that has taken a share of MATLAB's territory," with MATLAB six ranks and 0.13 points below it. Both halves of that are true at once. Julia is winning its category and its category is roughly one twenty-fourth the size of Python's.

The project's own annual survey drew 721 responses in 2025, down from a 2,660 peak in 2021, with no year since 2023 above 1,400. In fairness, 2025 was up on 2024's 682, so the line is a plateau rather than a slide. That instrument recruits through Julia's own Slack, Discourse, conference site and company site, so it measures existing enthusiasts rather than the market, and its quoted margin of error is decorative. The survey's top non-technical complaint, year after year, is that there are not enough Julia users in the respondent's own field.

There is also a startup cost the benchmarks hide. My first solve took 0.276 seconds after a 5.4 second script run. The same 60-job model in the warmed-up scaling run took 0.055 seconds. The difference is compilation, and it is the single most common complaint about Julia from people who have actually used it. Version 1.9 improved this dramatically, cutting time-to-first-plot for several major packages by one to two orders of magnitude against version 1.7, at the cost of 10 to 50 percent longer precompilation and larger cache files. It is better than its reputation and it is not gone.

And there is the concentration question, which cuts both ways and deserves to be asked out loud rather than buried. Half the top ten Julia contributors by commit count are on one company's payroll, by that company's own statement. If you like Julia, that is why it has full-time maintenance. If you are underwriting a ten-year product decision on it, it is a thing to know. The mitigations are real: the license is MIT and every version already published stays published under it, the registry is public, the governance charter names five people who are not all at one company, and every version already published stays published. The risk is not that someone takes it away. It is that the pace of a language you depend on tracks the commercial priorities of one firm.

Who this is actually for

Most small businesses do not simulate anything and never will, and an article that pretends otherwise is selling something.

The federal numbers set the scale. Census figures for the 2022 reference year put 235,088 of 239,265 US firms operating manufacturing establishments below 500 employees, 98.25 percent of them, and those firms account for 40.9 percent of manufacturing employment. The Bureau of Labor Statistics puts the median wage for an operations research analyst at 88,940 dollars as of May 2025, and there are about 110,000 of them in the entire country. Census's Business Trends and Outlook Survey, for the collection period ending 3 May 2026, put AI use at 19.8 percent nationally and 37 percent among firms with at least 250 employees, and declined to publish a point estimate for the smallest band at all, printing only "less than 20 percent."

So the honest audience is narrow. It is the shop with a real allocation problem: which jobs, which machines, which week, which driver, which shift. It is the energy or HVAC consultancy that models the same building forty times a year. It is the small pharmacometrics or biostatistics practice, which is the one vertical where the Julia story has a genuine commercial product with a track record. It is the operator currently paying per-seat for something used two weeks a year. And it is anyone whose analysis has outgrown a spreadsheet but does not justify a four-figure annual seat plus a prerequisite stack.

If that is not you, the answer is that this is a fascinating story about how software gets owned and you can safely go back to work. If it is you, the free layer will take you a great deal further than the pricing pages suggest, because the free layer is where the results that check out were actually produced. The NASA separation runs, the MERL refrigerant work, the Lincoln Laboratory work: the primary sources for those name the open packages, not the paid platform. Instron is the exception, and JuliaHub says so plainly, crediting that 500x to the commercial tool.

The ledger reading

Strip both stories down and they are the same story told from opposite ends.

The Air Force got a simulation framework it holds the maximum data rights in, which the regulations grant as a license rather than as ownership, and it cannot put it on the internet because a different body of law says no and nobody at AFRL can unilaterally say yes. Four researchers at MIT wrote a language, published it under a license that grants everything and asks for a copyright notice, and by publishing it rather than by licensing it put it outside the export regulations for good. The rules that hold AFSIM run the other way, and on that side of the line publishing is not by itself a route out. One of those is a story about ownership. Neither is a story about generosity.

What sits on top is where the money is, in both cases. AFSIM has three contract vehicles running to 2029 and a base-and-all-options figure of 85 million dollars against zero obligated so far. Julia has a company that announced a 65 million dollar Series B in April 2026, against a securities filing showing about 66.7 million banked including converted notes, employs half the top ten contributors, publishes 0-dollar, 20-dollar, 60-dollar and 200-dollar tiers that a business may use for thirty days, and stopped publishing its platform prices in the second half of 2025, about two years after it first put dollar figures on that page. The page had carried the heading "Complete Transparency, Full Control" since 2022, back when the same two tiers said only "Request Pricing."

The reason to check any of this is not that anyone lied. It is that the checkable numbers and the marketing numbers come apart in a specific, repeatable place, and that place is always the same: the moment a result leaves the document that measured it. Jonnie Diegelman said 15,000 and immediately said the number mattered less than the tool being fast, and named his baseline anyway. The tile that quotes him dropped the baseline, dropped the workload and changed the subject of the work. Nobody had to lie for that to happen. Somebody just had to write a shorter sentence.

The free thing runs on my machine and cost nothing and I can tell you exactly what it did, which is 60 jobs in 0.276 seconds and 100,000 in 133. That is the whole argument for the open half. It is not that it is better. It is that you can check it.

Fact-check notes and sources

The MIT News article, its publication date, and every claim attributed to it: "How an MIT research project became a global programming language", Zach Winn, MIT News, 31 August 2026. The nine-body-link count and the eight-of-nine JuliaHub share are my own count of anchors in the article body, excluding site chrome and share buttons, taken 6 September 2026.
The Julia license, the copyright years, and the grant language: LICENSE.md, JuliaLang/julia. The bundled copyleft dependencies: THIRDPARTY.md. Note that the MIT license is a copyright grant and is silent on export control; the two are separate questions and this article treats them separately.
"90% of our ungracious demands" and the retroactive footnote about the 1.0 wording: "Why We Created Julia", 14 February 2012. The 1.0 release date and the API-stability commitment: "Announcing the release of Julia 1.0", 8 August 2018.
The committers, the Stewards, the five NumFOCUS signatories, and the "Related Organizations With No Governance Role" section naming both the MIT Julia Lab and JuliaHub: julialang.org/governance, read 6 September 2026. "Half of the top ten contributors by commits" is Stefan Karpinski's own phrasing in JuliaHub's blog. Note that the post carrying it has been back-edited to say "JuliaHub" throughout, years before that name existed, so read it as evidence of the current relationship rather than the historical one.
14,284 packages in the General registry: my own count of path = entries in Registry.toml, 6 September 2026. The registry is append-only, so this is "ever registered and still listed," not a count of maintained packages. The "over 12,000" and "13K+" marketing figures are floor claims and are therefore true, if stale. The "1 million users" and "100 million downloads" figures appear only on JuliaHub marketing surfaces, with no published methodology; I could not verify either and neither should be repeated as measurement.
Distribution Statement F's current authorised wording: DoD Instruction 5230.24, effective 10 January 2023, paragraph 4.2.f, corroborated by DTIC's own marking guide. The canonical esd.whs.mil copy refuses non-browser clients; the text above was read from a mirror and cross-checked against the DTIC guide, which is on a .mil domain. Note also that the current instruction reads "Further distribution" while AFSIM's published marking still uses the superseded 2012 wording, "Further dissemination." Both are quoted here as their own source prints them.
The definition of unlimited rights: DFARS 252.227-7014(a), clause dated August 2025. Data rights and distribution rights are different locks, which is why AFSIM can be fully government-owned and still undownloadable.
The Boeing IR&D funding window and the February 2013 delivery under contract: the Boeing and AFRL co-authored 2015 conference paper says the framework was "developed and funded through internal research and development (IR&D) funding from 2003-2014" and delivered "under contract" with unlimited rights. It carries no dollar figure, and the widely repeated 35 million dollar number comes from a later CSIAC article rather than from the delivery record, so this piece does not print it. The "more than 275 organizations" figure is from RAND RR-A1722-6, which is the Department of the Air Force's own federally funded research center writing under Air Force contract, and the report gives no source for the count; treat it as an AFRL figure repeated, dated to no later than 2024.
AFSIM's distribution terms, the "government-owned, open-sourced" description, the Information Transfer Agreement route, and "All output from AFSIM is also export controlled, regardless of input": DSIAC's AFSIM page. That page is stamped "POSTED: October 4, 2023" with no visible revision date, so treat it as a late-2023 snapshot.
Published software being outside the scope of the export regulations: 15 CFR 734.3(b)(3) and 15 CFR 734.7(a). The ITAR requirement for prior government approval: 22 CFR 120.34(a)(7). Two limits on the carve-out worth stating: it is an export-regulations exclusion only and does nothing for items controlled exclusively under the ITAR, and published encryption software under ECCN 5D002 stays in scope under 734.7(b).
The SHARE IT Act: Public Law 118-187, approved 23 December 2024. The sharing duty is Section 3, not Section 4; Section 4 is the exemptions section. Section 3(a) accepts "not less than 1 public repository or private repository," so it is not an open-source mandate.
The three AFSIM FORGE awards, their period of performance, the zero obligations and the 85,000,000 base-and-all-options field: the USAspending API, queried 6 September 2026, for FA239126DB001 (Radiance Technologies), FA239126DB003 (Stellar Science) and FA239126DB004 (Infinity Labs). FA239126DB002 returns no record. The 85 million figure appears on each of the three awards, so whether it is one shared ceiling or three separate ones is not settled by the data. Zero obligated is normal for an indefinite-delivery vehicle before task orders post.
The aircraft collision claim: the page MIT News links is JuliaHub's Lincoln Laboratory case study, which contains no occurrence of "Python," no occurrence of "faster," and no speed multiple. I could not find a primary source for "about 50 times faster" anywhere. The underlying primary material is Robert Moss's 54-slide ACAS X deck from JuliaCon 2015, dated 29 June 2015 and sponsored by the FAA per its own title slide, which mentions MATLAB three times, Python zero times, and states only that exhaustive comparison over 6.5e11 decision points "would be computationally expensive for MATLAB." The word "faster" appears twice in that deck, both times describing the ACAS X program rather than the language.
The WhatsApp user count: Meta's last disclosed figure is more than 3 billion monthly users, from the company's first-quarter 2025 earnings call in May 2025. No source at any tier reports 4 billion. That sentence in MIT News carries no link.
The Moderna work and its disclosure: the peer-reviewed pharmacometrics paper states that two Pumas-AI authors "were paid by Moderna, Inc., to perform the analysis," on a Moderna- and BARDA-funded study with a Moderna corresponding author. The Comrade.jl "100x" figure is from a ten-minute JuliaCon 2022 lightning-talk abstract and does not appear in the peer-reviewed JOSS paper, which contains no speed claim at all and says only that the package "has already been widely used within the Event Horizon Telescope Collaboration."
The six homepage tiles: read from juliahub.com on 6 September 2026, and each compared against the page it links to. Instron's 500x, MERL's <2% and NASA's 15,000x all appear in their linked sources. ASML's 50%/700+, Boeing's two-versus-five-years and Williams' 50x/4x appear neither on the landing page each tile links to nor in the talk behind it. The ASML and Boeing tiles link to JuliaHub video pages, not to the conference talks. The MERL result is simulation against simulation; the customer's own paper reports the estimate "for this simulated case," and no physical unit was weighed.
The Instron units error: the case study's own inputs, "10+ minutes" per loss-function evaluation and "1000+ iterations," multiply to 166.67 hours, and the document prints "well over 166 days." The case study never defines the relationship between a system run and a loss-function evaluation, so the denominator of the 500x cannot be checked from the document. Williams Grand Prix Technologies' own hosted copy of its case study carries a different stat block (169x faster, 7% more accurate, 1000x speed, 2.3x fidelity), names JuliaSim eight times and Dyad zero times, and does not mention MATLAB, where JuliaHub's copy mentions MATLAB five times.
JuliaHub's 2023 price list, headed "Complete Transparency, Full Control": Internet Archive capture, 23 September 2023. The June 2025 repricing: capture of 22 June 2025. The archive holds no capture of that page between 22 June 2025 and 1 December 2025, and the first capture of the replacement lead-capture page is 6 October 2025, so the correct statement is that pricing disappeared between those two dates, not on a specific day.
I did not install or evaluate Dyad. Every statement about it in this article concerns its published pricing page, its license terms and its marketing material, and none of it is an assessment of the software. Where the piece compares the free stack to what Dyad and Simscape sell, it is comparing something I ran against a feature list I read. The route from Dyad to the EULA is the license pointer on the Dyad Studio listing, juliahub.com/company/eula, which returns a permanent redirect to /legal/eula; I followed it on 6 September 2026.
The current platform pricing page with no prices: juliahub.com/pricing/juliahub, read 6 September 2026. The Dyad tiers at $0, $20, $60 and $200 a month, and the "Non-Commercial Use" heading above them: juliahub.com/pricing/dyad, same date. The thirty-day commercial trial and the non-commercial-only period after it: the JuliaHub EULA, last modified 3 August 2026, quoted verbatim above. Do not read the published Dyad prices as prices a business may pay to use it commercially; they are not.
The benchmark paper and the funding announcement's characterization of it: the paper is "Improving process systems engineering with specialized multi-agent large language models", Chemical Engineering Journal Advances vol. 26, May 2026, article 101141, gold open access. Every quotation of the paper above is from the journal's own indexed abstract as served by the DOAJ record, because the publisher's site does not serve the body text to an ordinary reader; I have not read the full paper and nothing above is drawn from it. The four systems tested are Dyad 2.1 and Claude Opus 4.6 as multi-agent systems and ChatGPT 5.2 and Gemini Pro 3 as single-agent baselines, all at their early-March-2026 state. The "barely completed the initial setup" sentence, the naming of Codex, and both the "months to mere days" and "months to minutes" figures are quoted from JuliaHub's Series B announcement of 30 April 2026, read 6 September 2026. Author affiliations are from the paper's own author block; eight of thirteen carry Julia Computing. Funders are from the Crossref record: CAPES, FAPERJ and CNPq, with JuliaHub not listed. I could not read the competing-interests declaration, so this article makes no claim about it beyond the affiliations the author list states.
PolyForm Noncommercial 1.0.0 expressly permitting government-institution use: the license text. The Dyad Source Available License is derived from it and inverts that clause. Dyad's open component libraries are genuinely open, BSD-3 licensed, carrying both the Modelica Association's original copyright and JuliaHub's translation copyright, and BSD-3 did not compel that choice: it requires notice retention, not open relicensing. JuliaHub chose it.
Dyad is JuliaSim renamed. JuliaHub's own blog carries the phrase "Dyad (formerly JuliaSim)," and the Dyad changelog runs unbroken from 0.5.1 on 16 June 2025 through 1.0.0 on 14 August 2025, 2.0.0 on 20 December 2025, 3.0.0 on 13 May 2026 and 3.3.0 on 6 August 2026. Anything dated before mid-2025 describing JuliaSim is describing the same lineage.
Every MathWorks price quoted: read from mathworks.com/pricing-licensing.html rendered in a browser on 6 September 2026, and from the North America Standard perpetual price list dated March 2026. A plain fetch of that pricing page returns a pre-render default tile showing USD 940 regardless of which product is selected; the correct annual figures are 1,050 for MATLAB and 1,585 for Simulink and appear only in the rendered page. The 3,685-dollar Simscape stack total is my arithmetic on three published prices, not a published bundle. The Home license restriction sentence as quoted is from the MathWorks pricing page with intended use set to Home, rendered in a browser on 6 September 2026. The MATLAB Home product page states the same restriction in different words, "For personal use only. Not for government, academic, research, commercial, or other organizational use," so quote whichever page you actually opened rather than merging them.
Ansys, Dassault Systèmes and COMSOL publishing no price: checked individually on 6 September 2026. Siemens is the exception: its Amesim product page carries no price, but a named Simcenter Amesim configuration sells at 4,272 dollars with an add-to-cart button on Siemens' own storefront. Gurobi's pricing page carries no dollar figure and one pricing call to action. IBM does publish a CPLEX figure, "Starting at $320.00 USD per authorized user per month," for a subscription that "allows only development use," with deployment priced on request and a discount offer live on the page the day it was read.
OpenModelica's current release, its consortium's 2007 founding, and the express permission for commercial use: openmodelica.org and the OSMC bylaws. The license discrepancy is real and matters: the shipped OSMC-PL version 1.8 makes the non-member mode GNU Affero GPL v3, and the bylaws revision history records the replacement of GPL v3 with Affero, while the website's summary still describes GPL v3. The Modelica Standard Library's BSD-3 license and version 4.1.0, announced in May 2025 (the exact day appears in the announcement URL rather than in the page body): the Modelica Association's release announcement. The counts of 1,430 component models and blocks, 534 examples and 1,219 functions are from the library's own 4.1.0 documentation rather than that announcement.
ModelingToolkit's transitive licenses are not all MIT. Its own LICENSE.md is MIT and ModelingToolkitBase is MIT, but as of version 11 the index-reduction and tearing algorithms sit in StateSelection and ModelingToolkitTearing, both AGPL-3.0 and both hard dependencies, so Pkg.add("ModelingToolkit") installs AGPL code. I checked this in my own package depot after the install described above: both LICENSE files are the GNU Affero General Public License version 3, and both packages are listed in ModelingToolkit v11.41.0 deps. It matters only if you redistribute or host, and it matters a great deal then.
ModelingToolkitStandardLibrary.jl is MIT licensed, copyright Julia Computing: the LICENSE file in the SciML repository and the documentation.
The package-liveness count: my own, computed on 6 September 2026 by shallow-cloning JuliaRegistries/General and counting distinct packages with any version registered since 6 September 2025. It is repeatable and it is an upper bound, because a single patch release counts a package as live. The routing sentence is from JuMP's Traveling Salesperson tutorial, which presents both an iterative method that runs on HiGHS and a callback method that does not. PackageCompiler's app-bundle behavior, and the absence of any published bundle size, are from its documentation; the 1.1 MB trimmed executable and the experimental status of --trim are from the Julia 1.12 release announcement.
Solver licenses: HiGHS is MIT, JuMP is MPL-2.0, Clarabel and SCIP and OSQP are Apache-2.0, Cbc and Clp and Ipopt are EPL-2.0, GLPK is GPL. License identification is from the LICENSE files themselves; what any of those obligations means for a specific product you ship is a question for counsel, not for a blog post.
The solver benchmark figures: Mittelmann's mixed-integer benchmark dated 7 July 2026 and linear programming benchmark dated 1 July 2026. Those two pages use different time limits and different machines and should not be read as one run. Unsolved instances are charged the full time limit, which inflates the ratios. The SCIP figure is the all-free SCIP-plus-SoPlex build. Gurobi, CPLEX and Xpress are absent because their vendors asked to be removed or withdrew, per Mittelmann's own note on the benchmark index page, not because they performed badly. That note is on the index, not on either results page.
Every timing, file size and result in the "what a small business can actually run" section is my own measurement on one Windows machine on 6 September 2026, running Julia 1.12.7 with JuMP and HiGHS installed from a clean depot. The job data is synthetic, generated by a deterministic sequence written into the script so that anyone running it gets identical numbers. The 10.65 percent margin gap is arithmetic on invented jobs; the mechanism it demonstrates, that greedy selection starves shared capacity, is general, and the specific percentage is not.
The thermal model, its component values, the reduction to one unknown and one equation, the 17.4-hour time constant and the 3.87-hour crossing: my own run on 6 September 2026, ModelingToolkit v11.41.0 and ModelingToolkitStandardLibrary v2.29.7 on Julia 1.12.7. The component values are mine, not measured from a real building, so the numbers describe the model I specified rather than any actual shop; the point being demonstrated is that the tool derives the equation from the connections, not that this is anyone's building. The simulated curve and the closed-form time constant agree, which is a check on the arithmetic and not on the physics.
TIOBE's September 2026 figures and its editor's comments: the TIOBE index. The table reads Python at 1 with 17.76 percent, Julia at 21 with 0.74 percent and MATLAB at 27 with 0.61 percent. TIOBE counts search-engine results, not employment or installed base. Its headline that month asks whether Julia will re-enter the top 20 and credits it with taking share from MATLAB, so the pessimistic reading and the optimistic one are both in the same source.
The Julia community survey response counts (2,660 in 2021, 1,162 in 2022, 1,329 in 2023, 682 in 2024, 721 in 2025) are from each year's own published deck. Respondents are recruited through Julia's own community channels, so the instrument measures existing users and its published margin of error should not be read as a sampling statistic.
US manufacturing firm counts and employment shares: Census Bureau Statistics of U.S. Businesses, 2022 reference year, NAICS 31-33. Note that SUSB assigns a firm's size band from its total employment across all industries while reporting only the employment in that industry's establishments, so these are firms operating manufacturing establishments rather than firms whose whole payroll is in manufacturing. The operations research analyst wage and headcount: BLS, May 2025; the Occupational Outlook Handbook and the OEWS tables give 113,100 and 108,510 respectively, hence "about 110,000." The AI adoption figures: Census Business Trends and Outlook Survey, collection period ending 3 May 2026. Census printed "less than 20%" rather than a point estimate for the smallest firms, so no meaningful gap between micro firms and the 19.8 percent national rate can be read from it.

This post is informational and journalistic. I have no affiliation with, and no financial interest in, MIT, JuliaHub, MathWorks, Ansys, Siemens, Dassault Systèmes, COMSOL, the Open Source Modelica Consortium, or any AFSIM contractor, and nothing here is endorsed by any of them. Nothing in this post is legal advice; license obligations depend on facts a blog post cannot know. Prices, versions and page contents were read on 6 September 2026 and all of them change.

Julia: What MIT Actually Gave Away, and What JuliaHub Will Not Put a Price On