Open Models Briefing: The October Surge

Briefing current as of 7 November 2026. Benchmark figures cited here are a mix of independent measurements and developers' own claims; we flag which is which.

The open-weight AI ecosystem has had busy months before, but nothing quite like the five weeks from early October to early November 2026. In that window, an open model came within a whisker of the closed frontier, three Western developers shipped weights in the hundreds of billions of parameters, and the question of what "open" actually permits became impossible to ignore. Here is what happened and what it means.

The headline: Kimi K3.1 Pro closes in

On 23 October, Moonshot AI released Kimi K3.1 Pro, a heavily post-trained version of its July model Kimi K3. The architecture is enormous, around 2.8 trillion parameters in total with roughly 104 billion active per token, and the context window extends to a million tokens. API pricing was set at 2 dollars per million input tokens and 10 per million output, a third below its predecessor.

The significant number came from Artificial Analysis. On its v4.3 index, K3.1 Pro scored 55 at launch, the highest score yet recorded for an open-weight model and three points behind the overall leader, Anthropic's Claude Opus 5.5. Four days later, on 27 October, Artificial Analysis replaced that index with v5, in which 60 percent of the tests are private. On v5, K3.1 Pro scores 46 against Opus 5.5's 58, so the gap is 12 points rather than 3. It is still the highest-scoring open-weight model by a wide margin: the next best, DeepSeek V4.1 Flash and GLM-5.3, score in the low thirties.

Moonshot promised the weights by 4 November and delivered. They are published on Hugging Face in a compressed 4-bit format, under a custom "Kimi K3.1 Pro License" that, like its predecessor's, includes clauses governing use as a hosted service and attribution. Two further details are worth knowing. First, Moonshot deliberately removed cyber-related training data from the open release, accepting a measurable drop on its internal offensive-security benchmark rather than open-sourcing that capability. Second, no consumer hardware will run this model; it is a data-centre deployment by any measure.

The West answers

As recently as April 2026, the ten strongest open-weight models on Artificial Analysis's leaderboard all came from Chinese labs. October changed that, at least in part.

Reflection: Beam (weights 16 October). The US start-up announced Beam on 5 October and published the weights eleven days later. It is a sparse mixture-of-experts model with 501 billion total and 23 billion active parameters, text-only, with a one-million-token context. Reflection claims parity with Z.ai's GLM-5.2 on reasoning benchmarks at three to four times less inference compute; that claim is the company's own and has not been independently verified. What sets Beam apart is the licence: plain Apache 2.0, with no user caps, revenue clauses or attribution requirements, and an explicit statement that distillation and training on outputs are allowed. Dell and HPE are offering it for on-premises and air-gapped deployment.

Meta: Muse Spark Open (15 October). Released a day before Beam at 816 billion parameters, Meta's return to large open weights came with a more restrictive licence, including attribution requirements and terms for companies selling API access above a revenue threshold.

Mistral: Large 4 (weights 29 October). The Paris-based lab previewed its new flagship on 6 October and published the weights at the end of the month. It is natively multimodal with around a trillion parameters, and Mistral positions it as the strongest open-weight model developed outside China. The company's self-reported agentic coding score improved substantially between the preview and the final release, a reminder that early numbers for open releases can move. For European buyers, the fact that it was built and can be served from Europe is as significant as its benchmark position.

Nvidia's Nemotron line continues in the background, with its largest open model of the year released in June. Reported filings indicate Nvidia plans to spend tens of billions of dollars on open-weight development, a strong signal that at least one of the industry's most powerful companies sees open models as strategic rather than charitable.

Announced is not released

A recurring theme of the period is the gap between announcement and availability. Several notable models have been announced as "open" but cannot yet be downloaded:

• Thinking Machines' Inkling 2 Small (5 November) scored 31 on the Artificial Analysis index. Artificial Analysis says that, once the weights are released as planned, it would be the highest-scoring open-weight model developed outside China on its current rankings, which sits awkwardly alongside Mistral's own claim for Large 4 and is a good illustration of why independent numbers matter.
• Alibaba's Qwen4 has four named variants, but as of 6 November no model card, licence or downloadable weights had been confirmed. A smaller model described as an open-weight preview of the Qwen4 architecture has been reported but not yet verified.

The practical rule is simple. A model is open when the files and the licence are on a public repository and you have read the licence. Until then, it is a closed model with a promise attached.

The licence question gets serious

As open models grow more capable and more commercially valuable, licences are diverging. October's releases spanned the full range, from Beam's unrestricted Apache 2.0 through Meta's and Moonshot's custom terms. For individuals and researchers the differences rarely matter. For a company building a product, offering a hosted service or operating at scale, they can be decisive. Licence review now belongs in model selection alongside benchmarks and hardware requirements, not after them.

The gaps that remain

The headline narrowing should not obscure where open models still trail. In October, Epoch AI found the best open model six to seven months behind the closed frontier on cyber capabilities, wider than the four-month average gap it had measured across general capability. Long-horizon agentic tasks and factual calibration remain the other areas where closed systems hold a meaningful lead. And at the very top, open increasingly means "open to anyone with a GPU cluster": the most capable releases are far beyond anything a single workstation can host.

The sovereign and local angle

For organisations focused on sovereignty or local deployment, the most relevant developments are not always the headline giants.

Apertus continues to mature. Version 1.5, released in July with a 262k-token context window, a reasoning mode and image understanding, reached general availability in September and is now offered by a range of Swiss and European providers and in Proton's Lumo assistant. Training for Apertus 2.0, a much larger successor, is underway, with the first training phase scheduled to finish in December 2026 and a release planned for 2027.

Mid-sized models matter most in practice. Models in the 8 to 35 billion parameter range, the size that runs on a capable workstation or a single server, saw steady releases through the period; Nanbeige's 33-billion-parameter Nanbeige4.3, released on 5 November, matched Alibaba's similarly sized Qwen3.8 on the Artificial Analysis index. For local retrieval systems and departmental deployments, these are the models that will actually be used.

The regulatory backdrop

Two policy developments frame all of this for European and Swiss readers.

In the EU, the AI Omnibus amending the AI Act entered into force on 27 July 2026. It deferred most high-risk obligations to December 2027 and August 2028, but left the general-purpose model rules and the AI Office's enforcement powers on their original schedule, which took effect on 2 August 2026. Providers of general-purpose models, open-weight releases included, are now subject to active supervision.

In Switzerland, the Federal Council took note of interim progress on 28 October 2026. The approach remains to ratify the Council of Europe's AI Convention and adapt existing laws sector by sector, with a consultation draft now scheduled for spring 2027.

What to do with this

For anyone selecting models over the coming months:

1. Re-run your evaluations. If you settled on a model before October, the options have changed materially. Test the new releases on your own tasks rather than relying on leaderboards.
2. Read the licence in the repository, not the press release.
3. Separate the frontier from the deployable. The trillion-parameter models set the ceiling; the mid-sized models are what most organisations can actually run.
4. Wait for independent numbers. Several headline claims from October are still vendor-reported. Independent evaluations typically follow within weeks.
5. Watch December. Pending weight releases and the completion of Apertus 2.0's first training phase could reshape the picture again before year-end.