Open Weights vs Open Source (And Why an MIT License Is Not Enough)

Nearly every explainer on this topic tells you that open weights means you get the file and open source means you get the recipe, then stops. That framing is not wrong, but it teaches developers the wrong test — because a maximally permissive MIT license on a set of weights still does not make a model open source under the Open Source Initiative's definition. Here is the map we actually use, with every license checked at its primary source in July 2026.

Two axes, not one spectrum

The standard explainer treats openness as a single line running from closed API on the left to open source on the right, with open weights sitting somewhere in the middle. That mental model produces one very specific and very common mistake: a developer checks the Hugging Face license tag, sees mit, and concludes the model is open source.

It is not. License permissiveness and open-source-ness in the OSI sense are two independent axes. A model can be maximally permissive on the legal axis — MIT, no field-of-use restriction, no user cap, no attribution clause — and score near zero on the disclosure axis, because the training data, the data-filtering pipeline and the training code were never published. That is precisely where most of 2026's headline releases sit.

The Open Source Initiative's own community forum worked through this with DeepSeek-R1. The weights and the inference code were MIT. The thread's conclusion was still that R1 does not conform to the Open Source AI Definition, because, in the forum's words, the code to process and filter and train the data is missing — This is where the secret sauce is.

What open source actually means for a model

OSI published the Open Source AI Definition 1.0 on 28 October 2024, at the end of a year-long co-design process built on several years of prior research and workshops. It asks for four freedoms: use the system for any purpose without asking permission, study how it works and inspect its components, modify it for any purpose including changing its output, and share it for others to use with or without modifications.

To make those freedoms real, OSAID requires three components in their preferred form:

• Data information — sufficiently detailed information about the training data that a skilled person could build a substantially equivalent system.
• Code — the complete source code used to train and run the system.
• Parameters — the weights and other configuration settings.

All three must be available under OSI-approved terms. Weights alone do not get you there. OSI has a separate page for that case, where it defines open weights as the final weights and biases of a trained neural network, calls them merely a starting point and asks whether they are a lesser evil relative to proprietary AI, and lists the resulting gaps: lack of reproducibility, data opacity, insufficiency for emerging regulatory transparency mandates, and limited community collaboration.

OSI's FAQ names the models that passed validation — Pythia, OLMo, Amber, CrystalCoder and T5 — and the ones that did not: Llama 2, Grok, Phi-2 and Mixtral. OSI is careful to say these are not certifications of any kind. Note what is missing from the passing list: every widely deployed 2026 open-weight model. OSI has said it will keep evolving the definition as the field matures; commentators tracking the process expect a 1.1 or 2.0 revision that revisits the data-information compromise, though OSI has not published a firm date.

The five tiers you will actually meet

Most comparison tables have three columns: open source, restrictive, closed. In practice there are five distinct legal shapes, and the two everyone omits — the permissive-plus-rider tier and the purpose-built model-distribution license — are the ones that generate real procurement questions.

Licenses verified July 2026 at their primary sources (Hugging Face model cards, Moonshot's LICENSE file, llama.com, ai.google.dev, openmdw.ai). Licenses differ per model and per version — check the specific repo before you deploy.

Tier 5 deserves a note, because it kills the laziest mental model in this space. Alibaba is the Apache-2.0 standard-bearer — Qwen3-Coder-Next ships with an apache-2.0 tag, 80B total and 3B activated parameters, and a native 262,144-token context — and the same company ships its Qwen3.x-Max flagships as closed, API-only, with no downloadable weights. One vendor occupies both ends. Open versus closed is not a map of labs or countries; it is a per-release business decision. If you want the geopolitics angle specifically, we cover it separately in export controls and open weights.

Riders: Modified MIT, community licenses, remote kill switches

Kimi K2.7-Code carries the Hugging Face license tag spelled modified-mit — not mit. SPDX tooling will not recognise it as MIT, which matters if your compliance pipeline auto-classifies dependencies. Moonshot's text is standard MIT with exactly one addition: if the software or a derivative is used in a commercial product or service exceeding 100 million monthly active users or 20 million US dollars in monthly revenue, you must prominently display Kimi K2 in that product's interface.

The threshold is not the problem — almost nobody reading this will cross it. The ambiguity is. The license does not define font size, placement, or exact wording for that display. If you are the rare company that does cross it, you inherit an obligation with no objective compliance test.

Tier 3 is a different species of risk. Under the Llama Community License, an entity with more than 700 million monthly active users in the preceding calendar month must request a licence from Meta, which Meta may grant in its sole discretion, and an Acceptable Use Policy is incorporated by reference. Google's Gemma Terms of Use go further in a direction that surprises people: you must include the use restrictions as an enforceable provision in any agreement governing use or distribution of Gemma or model derivatives and notify subsequent users, and Google reserves the right to restrict — remotely or otherwise — usage of any Gemma Services it reasonably believes violate the agreement. A remote-restriction clause is the thing a self-hosting strategy is usually meant to eliminate.

The practical consequence, flagged repeatedly in legal analysis of these releases: weights-only distribution prevents you from verifying whether copyrighted works were used in training or identifying inherent biases, which is an auditability and IP-liability exposure, not a philosophical one. The same analyses warn about open washing — releasing partial components while claiming full open-source status.

Does open weights cost you quality?

Here is where most articles either overclaim or dodge. We ran the test. DataLLM Lab's first-party executed benchmark in July 2026 put 13 models through 9 generate-code-then-run-hidden-tests tasks — the model writes code, we execute it against tests it never sees, and cost is real token usage priced at list, scaled to 1,000 tasks.

DataLLM Lab first-party executed benchmark, July 2026. Cost is real token usage at list prices as of July 2026 — prices move, and self-hosting economics (GPU hours, utilisation) are a different calculation entirely from per-token API price. Full methodology in the coding cost benchmark.

Ten of the thirteen models scored a perfect 9/9. Correctness on this class of work is table stakes. The only variable that separated them was an 88x cost spread — Qwen3 Coder Next matched GPT-5.5's score at roughly one eighty-eighth of the price.

Now the part the hype pieces leave out. Do not generalise that into open weights have caught the frontier, because three independent trackers say otherwise. Epoch AI measures open-weight models lagging state-of-the-art closed models by about four months, roughly 8 points on its Capabilities Index, over January to May 2026 — widened from a three-month lag in its October 2025 estimate, and Epoch notes the gap would grow to six months under a stricter test. Stanford HAI's 2026 AI Index puts the top-closed to top-open gap at 3.3% as of March 2026, widened from 0.5% in August 2024. Artificial Analysis's 30 April 2026 roundup shows the best open models around 6 Intelligence Index points behind GPT-5.5, with much larger gaps on hard evals: 34-36% versus 44% on Humanity's Last Exam, 4-12% versus 27% on CritPt physics, 43-46% versus 61% on TerminalBench Hard agentic coding.

Both things are true, and the scoped version is the useful one: open weights are not a quality compromise for the bread-and-butter work most teams actually ship, at a fraction of the cost — and still are a compromise on frontier-hard reasoning. That is a routing rule, not a slogan.

One live example of how fast this moves: Kimi K3, a 2.8T-parameter MoE, reached #3 on Artificial Analysis's Intelligence Index and #1 on Arena's Frontend Code Arena. We ran it via API on its launch day (16-17 July 2026). It passed all 8 tasks we could execute; the 9th could not be evaluated, because parse_csv_line never completed under repeated upstream 429 rate-limiting across both runs. That is a launch-day capacity signal, not a capability miss. Real cost landed around $4 per 1,000 tasks, Opus-4.8 territory, and it was notably token-lean at roughly 150 reasoning tokens per task. Its weights were still not public as of 21 July 2026 — promised for 27 July, with no final license text published. Treat any Modified MIT expectation for K3 as extrapolation from the K2 family, not a confirmed term. We keep the details current in the Kimi K3 review.

Where the label has cash value

The open weights versus open source distinction is usually argued as a philosophy question. In the EU it has a compliance consequence. Article 53(2) of the AI Act exempts general-purpose AI models from the Article 53(1)(a)-(b) technical-documentation obligations, but only if three conditions all hold at once:

1. The model is released under a free and open-source licence permitting access, use, modification and distribution.
2. The parameters (including weights), model architecture information and model usage information are made publicly available.
3. The model is not provided against a price or otherwise monetised.

Read the tiers against that. Permissive software licences such as Apache-2.0 and MIT — and permissive model licences such as OpenMDW — are generally read by expert commentators at Hugging Face and Linux Foundation Europe as falling inside the free and open-source concept. A licence with a discretionary approval gate above a user threshold, mandatory flow-down of use restrictions, and a reserved right to restrict usage remotely is a much harder argument. That is a structural difference between tier 1 and tier 3 that nobody surfaces when they write these comparisons.

Two caveats that matter more than the exemption itself. First, the exemption is partial: the copyright policy obligation and the sufficiently detailed summary of training content survive regardless. Second, it evaporates entirely for models classified as posing systemic risk — training compute at or above 10^25 FLOP — where the full obligation set returns. And to be explicit: whether any specific licence qualifies has not been settled by a regulator decision or case law. This is a description of the statutory text, not legal advice; talk to counsel before relying on it.

A decision rule you can apply today

Skip the taxonomy debate and ask four questions in order.

1. Do you need to run it on your own hardware or in your own VPC? If yes, you need downloadable weights — tiers 1 through 4. This is also the only clean answer to data-residency concerns: self-hosting removes the data-transfer question entirely, whoever trained the model. Our local deployment guide covers what the hardware actually costs.
2. Will you redistribute a derivative, or ship it inside a product at scale? Then read the rider, not the tag. Tier 1 is clean. Tier 2 adds a display obligation with an undefined standard. Tier 3 can require a discretionary licence from the vendor once you get big enough.
3. Do you need to audit what the model was trained on? Then none of tiers 1 through 3 help — only OSAID-conforming releases like OLMo or Pythia give you data information and full training code. Every MIT and Apache-2.0 flagship in this article fails that test.
4. Is this task frontier-hard, or is it everyday? Everyday codegen, extraction, summarisation and refactoring: route it to open weights and pocket the 20x to 88x cost difference. Deep research, novel reasoning, long-horizon agentic work: keep it on frontier models until the independent trackers say otherwise. Our 2026 open-weight model roundup and the GLM-5.2 review break the per-model tradeoffs down further.

And one habit worth building: verify the licence at the model's own repository, every time, for the specific version you are deploying. Vendors ship different licences for different models in the same family, and the same family changes licences between releases.

FAQ