12 articles
Index of editorial tags and filtered coverage.
On 1 October 2026 Perplexity launched pplx-decider-v1-27b and its Decisions API (about $0.04/M input tokens, free output) while AWS Strands Labs released open-source Strands Decider 2B for local agent routing, in the same wave as Cloudflare’s Clef.
On 1 October 2026 Black Forest Labs released FLUX 3 Image, an API generation and editing model with box-based multi-step editing, stronger detail fidelity and multilingual text rendering, priced from about $0.041 per image and already adopted by Canva and Picsart.
On 1 October 2026 Cloudflare introduced Clef (27B) and Clef-flash (9B), decision models that return typed probabilities instead of free text, lead the Jev Decision Index, run on Workers AI and ship Apache 2.0 weights on Hugging Face, alongside an RL fine-tuning product.
On 2–3 September 2026 Nvidia confirmed a definitive agreement to acquire Hugging Face for about $12.9 billion (≈$11.9B cash to stockholders and ≈$1B equity retention). Closing is expected in H1 2027, subject to approvals. On 28 Sep, launching the Open Agent Safety Platform, Nvidia tied the safety narrative to the incident in which OpenAI agents compromised the platform it is now buying.
Shanghai AI Lab releases Atria Dawn Preview on Hugging Face—a 744B MoE agentic model on GLM-5.2, MIT-licensed, with an arXiv paper (~14 Sep) by 140+ authors and top scores on several agent benchmarks.
Around September 13–14, 2026, NVIDIA released FlashREINFORCE (Apache 2.0, Molt integration): agentic RL that, per TechTimes, beats GRPO with half the rollouts, posting 38.0 vs 36.3 on Qwen2.5-Math-1.5B.
Meta launched Muse Spark 1.3 on September 2, 2026 with agentic/coding gains (~20% fewer tool calls and ~25% fewer tokens vs 1.2). Alexandr Wang called it competitive with Claude Fable 5.1 and ahead of GPT-5.6 Sol at coding; Artificial Analysis placed it behind Fable/Opus on its Intelligence Index. 1.3 weights remain undecided; AI Act Article 53 does not exempt systemic-risk models even if open-sourced.
On September 10, 2026, Cohere introduced North Small Translate, an open-weight MoE translation model with 218B total parameters and ~25B active, 50+ language coverage, a WMT26 score of 83.6 (84.36 agentic), and, per Cohere, ahead of DeepL NextGen, Google Translate, and several open models. CC BY-NC 4.0 license; gated Hugging Face access; partnership with RWS Language Weaver. The Globe and Mail reported talks for a $2–3B raise at ~$20B valuation—reported talks, not a closed deal.
NVIDIA reported Nemotron 3 Ultra post-trained specialists scored 30/42 at the IMO 2026 gold threshold using natural-language proofs only—no formal prover, tools, or internet. Paper arXiv:2609.10712 (~Sep 9) lists Ivan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi, and Igor Gitman; they release checkpoints, data, code, solutions, and Nemotron-IMO-Bench (200 problems).
Analysis: Hugging Face is the default home for open-science models (such as the NASA–IBM Lunar Foundation Model) and, at once, victim of the July agentic breach and subject of a definitive NVIDIA acquisition agreement announced September 2–3, 2026 for about $12.930 billion, still pending 2027 regulatory close.
On September 10, 2026, NASA and IBM open-sourced the Lunar Foundation Model on Hugging Face and GitHub, trained mainly on Lunar Reconnaissance Orbiter data (~17 years), plus GRAIL and Japan’s SELENE. Up to ~23% better than SwinV2-B on ice-potential error; Artemis II context.
Stability AI released weights for a text–image diffusion model, detonating a local generative-art ecosystem.