llama.cpp release b11870 (25 Oct) added support for DeepSeek V4.1. The 17 Oct build has the joint-highest score among downloadable models on the Artificial Analysis Intelligence Index v5 (32), with vision, a 1M-token context and an MIT licence.
ISD view: why it matters
One of the two highest-scoring open models no longer needs special builds to run locally. Running it in-house sends no prompts or data to the lab, but check any US federal contract terms first.
Moonshot released Kimi K3.1 Pro through its API on 24 Oct and has promised downloadable weights by 4 Nov. On the new Artificial Analysis Intelligence Index v5 it scores 46; the highest-scoring models you can download today score 32.
ISD view: why it matters
If the weights arrive on time, this would be the highest-scoring model a business can run on its own hardware. Read the licence before planning on it: the terms are not yet published.
Recursive, Richard Socher's US lab, released Flint on 21 Oct under Apache-2.0: 39.8B parameters in total, about 2.5B active, 25.1 GB at 4-bit. Its benchmark scores are vendor-reported only so far.
ISD view: why it matters
Cheap to run on modest hardware, from a US lab, under Apache-2.0. On benchmarks both report, it scores below Qwen3.8-27B on accuracy; its advantage is speed and memory.
Meta (Muse Spark Open), NVIDIA (Nemotron 3.5) and Reflection AI (Beam 1.0) all released open-weight models in mid-October. The highest-scoring of them, Muse Spark Open, scores 25 on the Artificial Analysis Index v5; the highest-scoring open models overall, both from Chinese labs, score 32.
ISD view: why it matters
More choice for businesses that need to consider model origin, especially at the small, fast end, which is also the cheapest to run on your own hardware.
