← Daily LLM news

Daily LLM news, Thursday, 5 November 2026

Daily LLM Briefing: Thursday 5 November 2026

Kimi K3.1 Pro logo
Logo: Kimi K3.1 Pro (kimi.com)

Good afternoon. Moonshot has kept its promise and published the full open weights for Kimi K3.1 Pro; a new US lab has released a 65-billion-parameter coding model that tops other American open models on a major agent benchmark; a new cloud-ops benchmark finds reasoning, not knowledge, is where agents fail; Anthropic is expanding text watermarks on older Claude models; and NVIDIA researchers show a way to serve hybrid language models about a fifth to two-fifths faster without changing their scores.

LEAD: Moonshot publishes Kimi K3.1 Pro open weights on Hugging Face

Background: Moonshot launched Kimi K3.1 Pro in its apps and API on 23 October and promised the full weights "by Nov 4". Yesterday's briefing noted they were still missing at lunchtime UK time.

What happened: Late on 4 November (about 16:00 UTC, per Moonshot's X post and trackers that checked the repo) Moonshot published an ungated Hugging Face repository, moonshotai/Kimi-K3.1-Pro. It holds 96 safetensors shards totalling about 1.56 TB — native MXFP4 weights with MXFP8 activations from quantisation-aware training, not a later compression of a larger BF16 dump that some early coverage had assumed would be about 5.6 TB. Specs match the API model: 2.8 trillion total parameters, 104 billion active per token, 896 experts with 16 selected plus 2 shared, a 1,048,576-token context, and text plus image input. The architecture change from Kimi K3 is that full-attention layers become Gated DeepSeek Sparse Attention for cheaper long context (69 KDA and 24 Gated DSA layers). The licence is the custom Kimi K3.1-Pro License, not MIT or Apache: free use, modification and redistribution for most users, but Model-as-a-Service businesses with more than $20 million revenue in any 12 months need a separate Moonshot agreement, and products above 100 million monthly users or $20 million monthly revenue must display "Kimi K3.1-Pro". Internal use is exempt. Moonshot recommends vLLM, SGLang or TokenSpeed. ArtificialWatch ranks it as the top model with downloadable weights on its index. Running it still needs a datacentre-class cluster (vLLM recipes start around eight high-end GPUs); for many teams the API at $2 / $10 per million input / output tokens remains cheaper. A separate "Highspeed" variant has no public weights yet.

Why it matters: One of the strongest open Chinese frontier models is now inspectable and fine-tuneable, not only rentable — with a licence gate aimed at competing API hosts rather than product builders.

Sources: Hugging Face model card, ArtificialWatch, The Model Gap, OrcaRouter, GIGAZINE

Ricursive releases Ripple-65B Preview, claiming US open-weight lead on agentic coding

Background: Most recent US open coding models have been huge mixture-of-experts systems with hundreds of billions of parameters. Ricursive Intelligence is a small full-stack lab building models for its own chip-design platform.

What happened: On 4 November Ricursive published Ripple-65B Preview under Apache 2.0 on Hugging Face: a dense 65-billion-parameter Transformer, pretrained on 12 trillion tokens, with up to 1 million tokens of context. The company says it scores 80.52 on Terminal-Bench 2.1 — ahead of much larger US open models including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8) and NVIDIA Nemotron 3 Ultra 550B (56.4), all as reported on Ricursive's model card. It claims the full R&D cost from scratch was about $32 million, and that reinforcement learning reached that Terminal-Bench score with only about 1.4 million rollouts. Hosted access is listed on Together AI, GMI Cloud and OpenRouter, with Fireworks said to follow. The company stresses that many agentic scores are hard to replicate across harnesses, and some comparison numbers are its own evals.

Why it matters: A smaller, freely licensed US open model that claims to beat far larger peers on coding agents is a notable counterpoint to the "bigger MoE wins" pattern — though independent leaderboard confirmation is still thin.

Sources: Ricursive, Hugging Face

CloudOpsBench: agents fail cloud tasks mostly on reasoning, not missing knowledge

Background: Coding benchmarks such as SWE-bench test software fixes; cloud operations — repairing services, rotating keys, migrating workloads, decommissioning projects — are harder to grade and more safety-critical.

What happened: A University of Michigan team led by Ang Chen released CloudOpsBench v1.0: 45 AWS DevOps tasks that run the same way in real AWS and in a high-fidelity emulator. On the public leaderboard, Claude Fable 5.1 (max, Claude Code harness) leads at about 60.6% resolution, GPT-6 Astra (max, Codex) at about 58.9%, with Gemini 4 Argon, Grok 4.7 and others behind; no model clears the set. In the team's failure analysis, knowledge gaps were rare — models knew the relevant AWS services — but about 80% of failures were reasoning mistakes (narrow surveys of the environment, incomplete scope, stopping early) and about 20% were wrong actions, some irreversible (for example aborting an in-progress multipart upload and destroying data). The project is Apache 2.0 on GitHub; GCP, Azure and OCI extensions are planned.

Why it matters: It is a rare public measure of frontier models on real infrastructure work, and it points the bottleneck at planning and thoroughness rather than missing docs.

Sources: CloudOpsBench, failure taxonomy, The Decoder

Anthropic expands Claude text watermarks; older models keep a sampling loophole

Background: Under the EU AI Act's transparency rules, Anthropic watermarks text from newer Claude models and is retrofitting older ones. The mark is an imperceptible pattern in word choice (a SynthID-style method); detection stays in private preview for eligible organisations.

What happened: Reporting based on Anthropic's email to API customers says a second retrofit wave now targets 16 November for Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 (Opus 4.6 had been expected around 3 November). Anthropic's own docs list those models among watermark-supported releases and warn that on Claude Sonnet 4.6 — and, per The Decoder's read of the email, the other models in this wave — the watermark may not apply if temperature, top_p or top_k is set away from defaults. Teams that rely on the mark for compliance are told to omit those parameters. Marking is meant to work across Anthropic's first-party products and cloud partners; customers cannot turn it off on supported models.

Why it matters: Watermark coverage is broadening ahead of EU deadlines, but a one-parameter sampling change can still leave unmarked text on several production models — a practical limit for anyone treating the mark as a compliance control.

Sources: Anthropic docs, Claude Help Center, The Decoder

NVIDIA's SEED makes hybrid LLMs cheaper to serve without retraining

Background: Hybrid language models mix a few full-attention layers with many linear-attention or state-space layers. Serving them spends a lot of compute writing memories during the prompt and a lot of GPU memory bandwidth rewriting large recurrent states during generation.

What happened: An NVIDIA paper posted around 3–4 November, Shallow Encoding Emits Deep Memories (SEED), finds two regularities in production hybrids (including Qwen3.8-Flash-Next, Kimi-Linear, GLM-5.3-Flash and Nemotron 3.5 variants): writing deep memories needs less computation than reading them for the next token, and most energy in recurrent states sits in a "state sink" plus a low-rank working memory. SEED freezes the original weights, runs most prompt tokens through fewer layers while "emitters" write the deep memories, and stores recurrent states in a compact form. On official exams across six checkpoints it is nearly lossless, and on an agentic-coding serving trace with SGLang on GB300 GPUs it raises Qwen3.8-Flash-Next throughput per GPU by about 1.2–1.4×. Code and training components are promised for release.

Why it matters: Hybrid open models are spreading; a training-free serving trick that keeps scores and cuts cost is immediately useful to anyone hosting them.

Sources: arXiv:2611.06084

Also worth a look

  • OpenAI Images 2.6: ChatGPT and the API gained GPT-Image-2.6 Flare (fast) and Sunburst (higher quality), GIF generation, markup edits and SVG export; OpenAI says people now create about 4 billion images a week across ChatGPT, Codex and the API. Image models, not text LLMs, but a notable OpenAI API change on 4 November. OpenAI, The Verge
  • Tencent Hy4 weights still missing: Yesterday's full Hy4 launch promised downloadable weights and a system card "in the coming days"; as of this briefing they have not appeared. Tencent
  • Claude Fable 5.5 / GPT-6.2 Astra: Still no official confirmation or catalog listing from Anthropic or OpenAI; treat as rumour only.

← Wednesday, 4 November 2026 Friday, 6 November 2026 →

Get the weekly sift

The best new AI and agent releases, with one-line verdicts. Occasional picks, no spam.

Privacy policy: your email goes to our email provider, Resend. Unsubscribe any time.

News is editorial and never paid for. Sponsored content on ToolSift is always labelled.