ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›BLOG›OPEN SOURCE VS OPEN WEIGHTS: WHAT IS THE
BLOG · COMPARISONBEGINNER

Open Source vs Open Weights: What Is the Difference?

Open weights means you can download and run a model’s trained files. Open source also shares the training code and data details, with no limits on use.

READ9 min
WORDS1,713
SECTIONS8
TYPECOMPARISON
CHECKED19 OCT 26

Open weights means a model’s trained files are published, so you can download it, run it and adapt it. Open source means more: you also get the training code and enough detail about the training data to rebuild it, under terms that let anyone use it for any purpose. Most AI models you can download are open weights, not open source.

TL;DR — THE SHORT VERSION
  • Open weights: you get the finished model file. You can run it and change it, within its licence.
  • Open source: you also get the code and the data details needed to rebuild it, and no limits on who uses it or for what.
  • The licence decides what you may do. Some open-weights licences cap company size, bar business use or leave out whole countries.
◈ IN PLAIN TERMS

Think of a cake. Open weights hand you the finished cake to take home, slice and decorate. Open source hands you the cake, the recipe and the shopping list, and lets anyone bake and sell their own. With open weights alone you can enjoy the cake, but you cannot bake the same one again from scratch.

Closed, open weights and open source at a glance
WordWhat you getCan you rebuild it?Limits on use
ClosedYou get: access through a website, app or API. The file stays with the company.Rebuild: no.Limits: the company’s terms, which it can change.
Open weightsYou get: the trained model file to download and run.Rebuild: usually not; the code and data details are usually not released.Limits: whatever its licence says. Some allow any use; some do not.
Open sourceYou get: the file, the training code and detailed information about the training data.Rebuild: a skilled person could build a substantially equivalent system.Limits: none on who or what for: use, study, modify and share for any purpose.

What does open weights mean?

The plain point: the company puts the model itself online for anyone to download, instead of keeping it on its own servers.

The weights are the model. They are the millions or billions of numbers an AI learns during training, and together they decide what it says. A US government agency describes them as values “the model has learned during training”.1 Once you have the file, a free program can run it on your own computer, and your chats stay there. Open weights meaning covers the basics and run AI on your own computer shows how.

The file is yours to keep, but the rules come from the licence published next to it. Some licences are standard permissive ones, such as MIT or Apache 2.0, that allow business use.2 Others write their own terms: Mistral’s licence for Mistral Medium 3.5 gives no rights at all to companies above a revenue line.3 AI labs outside the US compares more of these licences.

What does open source mean for AI?

The plain point: open source is a published standard, not a marketing word, and it asks for more than the file.

The Open Source Initiative (OSI), the group that keeps the definition of open source software, published an Open Source AI Definition. It starts with four freedoms: to use the system “for any purpose and without having to ask for permission”, to study how it works, to modify it, and to share it.4

To make those freedoms real, three things must be released. The model parameters (the weights). The code: “The complete source code used to train and run the system.” And data information: details about the training data good enough that “a skilled person can build a substantially equivalent system.” The definition says plainly that even “Open Source weights” must come with the data information and code used to make them.4

Why is open weights not open source?

The plain point: two gaps. Ingredients are missing, and many licences put limits on use.

Missing ingredients. In this site’s reading of the model pages it covers, most downloadable models ship the weights only. Without the training code and the data information, you can run and adapt the model but not rebuild it or check what went into it.

Limits on use. The OSI’s long-standing definition of open source software says a licence “must not discriminate against any person or group of persons” and must not stop anyone using the program “in a specific field of endeavor”.5 A licence that shuts out large companies, bars business use or leaves out countries fails that test. MiniMax’s licence for its H3 video model, for example, excludes “the European Union, the United Kingdom, the Republic of Korea and the United States of America”.6

A permissive licence closes only the second gap. Under the AI definition, an Apache 2.0 model still needs its code and data information published to count as open source.

Which AI models are actually open source?

The plain point: a few research models passed the OSI’s check; several famous downloadable models did not.

When the OSI tested its definition in 2024, its board report listed Pythia (EleutherAI), OLMo (AI2), Amber and CrystalCoder (LLM360) and T5 (Google) as passing. It said “Systems such as Llama2, Grok, Phi-2, and Mixtral did not meet the required criteria.”7 The OSI later said of Meta’s newer models: “Llama 3.x is still not Open Source by any stretch of the imagination.”8 That list is a 2024 snapshot. For any newer model, check the licence and whether the code and data information were released.

Why do companies call open-weights models open source?

The plain point: the word is used loosely, and it sounds better. The companies themselves blur it.

Xiaomi announced in September 2026 that it had “fully open-sourced the weights and technical report” of two models.9 Meta described its Muse Glimmer release as “open sourcing the model weights under a permissive Apache 2.0 license”.10 Both phrases describe weights. Neither, on its own, says the training code and data information were published, and this page has not checked whether they were.

TAKEAWAY

When a model is called open, ask two questions: what does the licence let me do, and were the training code and data details published? The word alone answers neither.

Does the difference matter to you?

The plain point: for running a model at home, rarely. For a business or a researcher, often.

Running a model yourself: open weights is enough. You get the privacy and independence of a file on your own machine. Local or cloud weighs that choice.

Building a product: the licence matters more than the label. Someone should read it before you depend on the model.

Studying a model: open source shows how it was made and what data went in. Weights alone show only what it does.

Either way, a release cannot be undone. NTIA notes that “once weights have been widely released, it is difficult to ‘un-release’ them”.11 The trade-offs for safety and power are on open or closed.

Questions people ask

Is Llama open source?

Not by the OSI’s definition. Llama models can be downloaded, so they are open weights, but the OSI listed Llama 2 among systems that failed its check and said Llama 3.x is “still not Open Source” (sources above). Is Llama dead? covers what Meta releases now.

Can I use an open-weights model in my business?

It depends on the licence of that exact model. MIT and Apache 2.0 allow business use with conditions such as keeping the notice; others cap company revenue, need a separate deal or exclude countries, as the Mistral and MiniMax examples above show.

Can a company take back an open-weights model?

Not really. Copies already downloaded stay out there; NTIA calls released weights difficult to “un-release” (source above). The company can stop offering new downloads or change the terms for future versions.

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What do you get with an open-weights model?

  • The trained model file, to download, run and adapt within its licence
  • The training code and full data details, with no limits on use
  • Access through the company’s website only
Show the answer

Answer: The trained model file, to download, run and adapt within its licence. The page says open weights means the trained files are published, so you can run and change the model, within its licence.

2. What does the Open Source AI Definition ask for beyond the weights?

  • A paid licence for business use
  • The training code and detailed information about the training data
  • A promise that the model is safe
Show the answer

Answer: The training code and detailed information about the training data. The page explains that open source also needs the code used to train and run the system and data information good enough to build a substantially equivalent system.

3. Why does a licence that shuts out large companies fail the open source test?

  • Because open source models must be small
  • Because only universities may publish open source
  • Because open source licences must not discriminate against people or fields of use
Show the answer

Answer: Because open source licences must not discriminate against people or fields of use. The page quotes the Open Source Definition: a licence must not discriminate against any person or group, or restrict a field of endeavor.

SOURCES
  1. 1NTIA (US Department of Commerce), Dual-Use Foundation Models with Widely Available Model Weights: Background, report published July 2024, read at source 1 Oct 2026 (as cited on this site’s open weights page): “The weights represent numerical values that the model has learned during training to achieve an objective specified by the developers.”↑
  2. 2Mistral AI, Mistral Small 4 model card, read at source 18 Sep 2026 (as cited on this site’s AI labs outside the US page), describing its Apache 2.0 licence as “Open-source license for both commercial and non-commercial use”.↑
  3. 3Mistral AI, Mistral Medium 3.5 licence, read at source 18 Sep 2026 (as cited on this site’s open or closed page): “You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million”.↑
  4. 4Open Source Initiative, The Open Source AI Definition – 1.0, read at source 19 Oct 2026“Use the system for any purpose and without having to ask for permission.” “Study how the system works and inspect its components.” “Modify the system for any purpose, including to change its output.” “Share the system for others to use with or without modifications, for any purpose.” ↑aCode, “The complete source code used to train and run the system.”; Data Information, “Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system.”; “‘Open Source models’ and ‘Open Source weights’ must include the data information and code used to derive those parameters.” ↑b
  5. 5Open Source Initiative, The Open Source Definition, version 1.9, page last modified 16 Feb 2024, read at source 19 Oct 2026“The license must not discriminate against any person or group of persons.” “The license must not restrict anyone from making use of the program in a specific field of endeavor.” ↑
  6. 6MiniMax, MiniMax H3 Community License Agreement, read at source 18 Sep 2026 (as cited on this site’s AI labs outside the US page): its “Excluded Territories” mean “the European Union, the United Kingdom, the Republic of Korea and the United States of America”.↑
  7. 7Open Source Initiative, Final Board Report, undated page (it refers to the definition’s release on 28 Oct 2024), read at source 19 Oct 2026systems that “successfully passed the validation process” were Pythia (Eleuther AI), OLMo (AI2), Amber and CrystalCoder (LLM360) and T5 (Google); “Systems such as Llama2, Grok, Phi-2, and Mixtral did not meet the required criteria.” ↑
  8. 8Open Source Initiative, Meta’s LLaMa license is still not Open Source, 18 Feb 2025, read at source 12 Oct 2026 (as cited on this site’s open weights page): “Llama 3.x is still not Open Source by any stretch of the imagination.”↑
  9. 9Xiaomi, MiMo-V2.6 release note, 22 Sep 2026, read at source 1 Oct 2026 (as cited on this site’s AI labs outside the US page): “We have fully open-sourced the weights and technical report of the MiMo-V2.6-Pro and Flash models”.↑
  10. 10Meta, Introducing Muse Glimmer, 10 Aug 2026, read at source 16 Sep 2026 (as cited on this site’s is Llama dead page): Meta is “open sourcing the model weights under a permissive Apache 2.0 license”.↑
  11. 11NTIA, Background, report published July 2024, read at source 1 Oct 2026 (as cited on this site’s open weights page): “once weights have been widely released, it is difficult to ‘un-release’ them”.↑

Read at source 19 Oct 2026: Open Source Initiative, The Open Source AI Definition – 1.0; The Open Source Definition (version 1.9); Final Board Report.

Reused from this site’s pages, with the read dates recorded there: NTIA’s open model weights report (1 Oct 2026) and the OSI’s Llama licence post (12 Oct 2026), on open weights meaning; the Mistral Medium 3.5 licence, Mistral Small 4 model card, MiniMax H3 licence (18 Sep 2026) and Xiaomi’s MiMo-V2.6 note (1 Oct 2026), on AI labs outside the US; Meta’s Muse Glimmer post (16 Sep 2026), on is Llama dead?.

The cake comparison and the advice on when the difference matters are this site’s own ways of explaining, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE