#openweights

Mistral raised €3 billion at a €21 billion valuation, nearly double its last round. It plans 1 GW of European compute by 2030 and will host rivals' open weights.

The InternLM team released Intern-S2-397B under Apache 2.0 on 11-13 September 2026, a multimodal mixture-of-experts model aimed at scientific reasoning and long agent tasks; the FP8 weights are a 406.3 GB download and the recommended setup is a node

Y Combinator chief executive Garry Tan said on 10 and 11 September 2026 that he would do nothing about AI model distillation and that US open-weight labs should be free to learn from closed American frontier models, putting him at odds with a federal

Agnes AI released open weights for a 33-billion-parameter multimodal Agnes 3.0 Flash under Apache 2.0 on 11 September 2026, but its model card says these are an earlier preview checkpoint and that the leaderboard score circulating with them belongs t

Researchers at Shanghai AI Lab and Shanghai Jiao Tong University released NCP-ArchPreview, an open 8.9-billion-parameter language model that learns to predict the next 'concept' as well as the next word, and reached the final training loss of a compa

The M-A-P research collective released YuE2 on 9 September 2026, a 3.6-billion-parameter open song generator that first writes an editable melody-and-chord score and then renders the full song, available as a 7.3 GB download under a non-commercial li

NVIDIA researchers released the full training recipe, data, 1.12 TB checkpoints and submitted proofs behind a Nemotron system that scored 30 of 42 points at the 2026 International Mathematical Olympiad, above the gold cutoff, as marked by the olympia

Cognition released SWE-2 on 10 September 2026, a coding model post-trained from the openly published Kimi K3 that reaches near-frontier coding scores at 64% lower cost, and which starts editing code after 18 exploratory steps where its predecessor to

DeepSeek released V4.1 Flash under an MIT licence on 10 September 2026: a 552-billion-parameter backbone plus a separate 196-billion-parameter memory module, split into an encoder and a decoder so that reading text costs half as much compute as writi

A project called Deltafin runs the full uncompressed Kimi K3 model — 2.8 trillion parameters, with 1.45 TB of expert weights — on a single MacBook Pro by streaming experts from four external SSDs on demand, sustaining almost exactly one token per sec

Tencent released AuK under an MIT licence, a 1.5-billion-parameter speech model that handles voice cloning, audio editing, enhancement and source separation through a single plain-language instruction interface, shipping as a 6.8 GB download with a d

A model identified as deepseek-v4.1-flash began answering on DeepSeek's API under an ID carrying its own expiry date, but the company has published no model card, no technical report and no changelog entry, and its predecessor V4 Pro remains the late