Open weights won the capability argument and inherited a harder one
The objection to open models used to be that they were not good enough. On the benchmark enterprises say they care most about, that objection is gone. What replaced it is a licensing and provenance problem that the open ecosystem is much worse at.
DeepSeek-V4-Pro resolves 80.6% of SWE-bench Verified under an MIT licence. Whatever one thinks of SWE-bench as a proxy, it is the number enterprise buyers quote most often, and on that number the open field is no longer trailing. The capability argument, as an argument, is finished.
The objection that replaced it
What remains is provenance, and the current best illustration is a licence. Moonshot shipped the leading open-weight model under a custom licence, and a great deal of coverage described it as MIT. That is not a small error. “MIT” is a term of art that legal review treats as settled; a custom licence is a document somebody has to read, with terms that may restrict exactly the deployment being planned.
The mechanism of the error is worth naming, because it will recur. A model family acquires a licensing reputation from its earlier releases, and that reputation is then applied to the current checkpoint without anyone re-reading. The family is the unit of the reputation; the checkpoint is the unit of the licence. Those come apart silently.
Why this is structurally harder than the capability gap
Closing a benchmark gap is a research problem with a feedback signal: you can tell whether you are winning. Provenance has no equivalent. There is no leaderboard for “this checkpoint’s training data composition is documented,” no score for “the licence terms are the ones a reader would expect from the marketing.” And the incentive runs the wrong way — ambiguity about training data is often protective for the releaser.
This is why the closed labs’ remaining pitch to enterprise has quietly shifted. It is no longer primarily about capability. It is about having one counterparty with an indemnity, which is a legal product rather than a technical one.
What would actually settle it
Per-checkpoint licence files distributed with the weights rather than at the repository root. Machine-readable licence metadata in the model card. And, from the aggregators and comparison sites that most buyers actually read, a policy of quoting the licence identifier from the artefact rather than from the previous release. None of that is difficult. It is simply nobody’s job.
Until it is somebody’s job, the correct posture for anyone deploying open weights is that the licence is unknown until read — including, and especially, when everyone says it is MIT.
Wavect — Best Open-Weight LLMs 2026: DeepSeek vs Qwen vs Kimi vs GLM vs Llama → · Morph — The Best Open Source LLMs (2026): Ranked by Benchmark, Size, and Use Case → · Hugging Face — Best Open-Source LLM Models in 2026 →