- More open-weight models to choose from, under a mix of licences.
- Artificial Analysis scores from before 27 Oct do not compare with today's.
- Firms with US federal contracts should review the DeepSeek order.
The month in brief
Meta, Reflection AI, NVIDIA and Recursive (all US) released Open-weight model: A model whose weights the lab publishes, so you can download it and run it on your own hardware. Glossary models in October, alongside new releases from DeepSeek and MiniMax (both China). The highest-scoring downloadable models remain DeepSeek V4.1-Flash-1017 and GLM-5.3 (both China, 32 on AA v5), with Muse Spark Open (Meta, US) at 25. Artificial Analysis moved to Intelligence Index v5 on 27 October and every score dropped, so older numbers no longer compare. And DeepSeek V4.1-Flash-1017 now runs in mainline llama.cpp.
Timeline
| Date | Event | Why it matters |
|---|---|---|
| 15 Oct | Meta Muse Spark Open; NVIDIA Nemotron 3.5 Super VL and Bolt | Highest-scoring open model from a US lab to date (AA v5: 25); NVIDIA local options |
| 16 Oct | Reflection Beam 1.0 (Apache-2.0) | A large model (US) under a permissive licence |
| 17 Oct | DeepSeek V4.1-Flash-1017 | Joint-highest score among downloadable models (AA v5: 32; China) |
| 19 Oct | US DHS makes public its 24 Jul order taking DeepSeek off federal civilian systems | Relevant to organisations with US federal contracts |
| 19-24 Oct | MiniMax-M3.1-Flash weights | AA v5: 29, but a restrictive licence |
| 21 Oct | Recursive Flint (Apache-2.0) | Small, fast coding model (US) |
| 24 Oct | Kimi K3.1 Pro by API; weights promised by 4 Nov | Highest AA v5 score among promised open weights (46; China) |
| 25 Oct | llama.cpp b11870 adds DeepSeek V4.1 | Runs in mainline llama.cpp without patched builds |
| 27 Oct | Artificial Analysis Intelligence Index v5 | All scores reset lower; closed models still score highest |
| 29 Oct | Mistral Large 4 weights not yet public | Overdue against the "end of October" promise |
What to watch in November
- Kimi K3.1 Pro weights and licence, promised by 4 November.
- Mistral Large 4 weights and licence.
- Independent scores for Recursive Flint and Muse Spark Open.
This month's items
- New scoring resets the Artificial Analysis leaderboard (27 Oct 2026)
- DeepSeek V4.1-Flash-1017 now runs in mainline llama.cpp (25 Oct 2026)
- Kimi K3.1 Pro: high scores by API, downloadable weights due by 4 Nov (24 Oct 2026)
- Recursive Flint: a small, fast coding model from a US lab (21 Oct 2026)
- US orders DeepSeek off federal civilian IT systems (20 Oct 2026)
- Meta, NVIDIA and Reflection AI release open-weight models in October (16 Oct 2026)
Earlier roundups: none yet (this is the first).