The best open-weight models in 2026, ranked

When this article first ran in July, it argued that the open-weight podium reshuffles monthly and that choosing on benchmark rank alone would leave an institution holding a stale decision. Eight weeks later, every name in the top three has changed. DeepSeek V4 Pro, then the most consistent all-rounder, no longer leads any major board. Zhipu's GLM-5.3 and Moonshot's Kimi K3 sit joint first among open-weight models on Artificial Analysis's index, at 44. GLM-5.3-Flash follows at 42, Qwen3.8 at 40, DeepSeek's V4 Pro 0813 at 36.

The scoreboard moved too, which is the more interesting development. Artificial Analysis has since revised its intelligence index to v4.3, so the six-point gap this article reported in July — measured on v4.1 — cannot be honestly compared with the nine points that separate the open-weight leaders from Claude Fable 5.1 and GPT-6 Astra today. The benchmark got harder; the models got better; the two effects are tangled together in a way no single number resolves.

That is not a reason to ignore the indices. It is a reason to rank on the things that hold still. What follows is ordered by deployability for an institution that runs its own stack — where licence terms, sizing and provenance carry as much weight as a score published this month and revised the next.

An August of releases

The summer was unusually busy. Z.AI shipped GLM-5.3 and, three days later, GLM-5.2 Turbo. Alibaba released Qwen3.8 27B under Apache 2.0 with native multimodality. DeepSeek put out a vision-capable V4 Flash variant alongside its MIT-licensed 304B Flash checkpoint. Mistral introduced Small 4 — 119B total parameters, 6.5B active, a 256K context window, Apache 2.0. On the proprietary side, OpenAI released GPT-6 Astra in early September, which is what a nine-point gap is currently measured against.

For an institution that wants models inside its own perimeter, the practical effect of that release cadence is not that the frontier moved. It is that the cost of waiting fell. A shortlist assembled in June was obsolete by September; a set of selection criteria assembled in June was not.

The ranking

1. GLM-5.3 family. Joint leader among open-weight models on the current index, MIT-licensed, with a Flash variant close behind the flagship and a Turbo release aimed at throughput. The family has been the most consistent riser of the past quarter, and the licence carries no revenue or user-count conditions — which keeps legal review short.
2. Kimi K3. Moonshot's line shares the top position and continues to be the strongest agentic performer of the group, which matters for multi-step workloads rather than single-response ones. The licence is Modified MIT, permissive below thresholds most institutions will never approach — but it is a clause your legal team should read rather than assume.
3. Qwen 3.8 family. The broadest catalogue on the list: many sizes under Apache 2.0, from a 27B multimodal checkpoint that runs on modest hardware to a 2.4T flagship scoring 40. Strong multilingual coverage keeps it on European shortlists. Breadth is the argument — one base-model family covering several use cases with a shared toolchain is worth more operationally than a single higher score.
4. DeepSeek V4 line. July's default, now mid-table on the indices and still an entirely reasonable choice: MIT-licensed, well understood, with a 304B Flash checkpoint and vision variants added in August. A model does not become worse because a scoreboard reorders. It becomes a less obvious purchase, which is a different thing.
5. Mistral. The European flagship, and the only name here operating under EU jurisdiction. Large 3 remains the 675B Apache 2.0 mixture-of-experts; Small 4 is the more interesting recent addition for deployment — 6.5B active parameters and a 256K context put frontier-adjacent behaviour on hardware a mid-sized institution already owns. On raw index scores Mistral trails the Chinese frontier, and provenance is what buys back the difference in procurement and vendor-risk review.
6. Llama 4 family. Scout's very long context remains genuinely unmatched for extreme long-document work. The custom licence, with its user-count cap and EU-specific restrictions, still ends the conversation for many European regulated institutions before the benchmarks are opened.
7. The small tier: Gemma and Phi. Apache-2.0 models that run on a single mid-range GPU. They score far below the flagships and sit closest to production for exactly that reason. A well-scoped small model doing one job well remains how most deployed work actually gets done.

What the scores still leave out

Three things a ranking compresses, all unchanged since July, which is rather the point.

Licences bite later. MIT and Apache 2.0 — GLM, DeepSeek, Qwen, Mistral, Gemma — permit commercial use without conditions. Modified MIT adds thresholds. Custom licences add geography. The cheapest moment to read a licence is before fine-tuning starts; the most expensive is after production dependency has formed.

Regulation attaches to use, not to weights. The Commission's enforcement powers over general-purpose AI obligations took effect in August, with fines of up to €15M or 3% of worldwide turnover for providers. Open-weight models retain partial exemptions where the model is genuinely open and outside the systemic-risk designation. The Digital Omnibus deferred the high-risk system obligations to December 2027 — deferral, not exemption, and the requirements still land on deployers whichever model sits underneath.

Benchmarks measure the general case; your documents are the specific one. A model scoring 44 and a model scoring 36 can trade places entirely on a portfolio of Polish loan agreements or DACH insurance policies. Frontier proprietary models keep their edge where work is genuinely open-ended — the honest case for the API route has not gone away. The gap that decides anything is measured on your own evaluation set, and no leaderboard publishes it.

Where this lands

Two months of churn have produced a useful natural experiment. An institution that selected on July's rankings would have chosen DeepSeek V4 Pro and would today own a mid-table model — running perfectly well, on a clean licence, doing the job it was scoped for. An institution that waited for the rankings to settle is still waiting, and has shipped nothing.

The podium changes monthly. Licence cleanliness, hardware fit, provenance a regulator accepts and performance on your own documents change on the timescale of years. Choose on the second list, revisit the first when a contract is up for renewal, and the reshuffles become news rather than a problem.

The list above is a starting grid. Which model wins depends on the race your institution is running — and that conversation is a shorter one than most expect.

Related reading:

• Open-weight AI models explained
• Open-weight models vs Claude and ChatGPT: the real decision criteria
• Governing open-weight AI models in regulated financial services
• How to choose a base model for your AI project

Benchmark sources: Artificial Analysis Intelligence Index v4.3 (September 2026) and v4.1 (July 2026); provider release announcements, August–September 2026. Scores as published at the time of writing and revised frequently.