Highlights
- Scores reset. Artificial Analysis moved to Intelligence Index v5 on 27 Oct and every score dropped: GLM-5.3 went from 45 to 32. New scoring resets the Artificial Analysis leaderboard
- Closed models still score highest. The top closed model, Claude Opus 5.5, scores 58; the highest-scoring downloadable models score 32. New scoring resets the Artificial Analysis leaderboard
- Easier to run in-house. llama.cpp release b11870 (25 Oct) added support for DeepSeek V4.1, one of the two joint-highest-scoring downloadable models. DeepSeek V4.1-Flash-1017 now runs in mainline llama.cpp
- One to watch. Kimi K3.1 Pro scores 46 through its API, with downloadable weights promised by 4 Nov and the licence not yet published. Kimi K3.1 Pro: high scores by API, downloadable weights due by 4 Nov
- Check which index version a vendor is quoting: v4 and v5 scores do not compare.
- Running a high-scoring model in-house is now simpler, but it still needs server-class hardware.
- Before planning on any new model, read the licence and check your contract obligations.
Scores: a new baseline
Artificial Analysis, an independent benchmarking firm, moved to version 5 of its Intelligence Index on 27 October, relying mostly on private tests. Scores quoted before that date do not compare with today's, so ask which version a vendor is citing. Read more: New scoring resets the Artificial Analysis leaderboard.
Running models yourself
DeepSeek V4.1-Flash-1017 now runs in mainline llama.cpp, so it no longer needs patched or third-party builds, which are harder to maintain and to trust. At about 500 GB for the Quantisation: Storing a model's numbers at lower precision (for example 4-bit) so it needs less memory. Glossary download, it is a server-class deployment. Running the weights locally sends no prompts or data to the lab that built them. Read more: DeepSeek V4.1-Flash-1017 now runs in mainline llama.cpp and AI hardware for on-premises use.
Licences and contracts
Kimi K3.1 Pro's licence is not yet published, so for most businesses the first question is the terms, not whether it can run. Organisations with US federal contracts should also check their obligations before choosing a model: the US has ordered DeepSeek off federal civilian IT systems. Read more: Kimi K3.1 Pro and US orders DeepSeek off federal civilian IT systems. Each model's origin is listed in the Open-weight model tracker so you can apply your own policy.
Coming up
- Kimi K3.1 Pro weights are promised by 4 November. Kimi K3.1 Pro
- Mistral promised the weights for Mistral Large 4 by the end of October; they were not public when we checked on 29 October. Kimi K3.1 Pro: also on the watch list
- For the whole month, see the October 2026 roundup.
Sources behind this week's items
- New scoring resets the Artificial Analysis leaderboard. Source: Artificial Analysis, “Announcing Artificial Analysis Intelligence Index v5”, 27 Oct 2026
- DeepSeek V4.1-Flash-1017 now runs in mainline llama.cpp. Source: llama.cpp (ggml-org, GitHub), “Release b11870”, 25 Oct 2026
- Kimi K3.1 Pro: high scores by API, downloadable weights due by 4 Nov. Source: Moonshot AI (Kimi), “Introducing Kimi K3.1 Pro”, 24 Oct 2026
- US orders DeepSeek off federal civilian IT systems. Sources: CNBC, “U.S. bans DeepSeek from government agency systems amid AI rift with China”, Jenny Lee and Kai Nicol-Schwarz, 20 Oct 2026; Congressional Research Service, “Removing PRC Artificial Intelligence Models from Federal Agencies”, 27 Oct 2026
Follow ISD Insights by Feed
Get this weekly overview, every new item and the monthly roundup in your feed reader.
FeedOther editions: none yet (this is the first).