DeAI Daily Brief — 6 November 2026

Today in DeAI: Moonshot AI's Kimi K3.1 Pro open weights are verified live, Thinking Machines Lab committed to full Apache 2.0 weights next week, and an EU court adviser said LLM training generally infringes copyright.

Kimi K3.1 Pro open weights are live, with a $20M MaaS clause

The promised weights for Moonshot AI's 2.8-trillion-parameter MoE are verified live and ungated on Hugging Face: 96 safetensors shards, MXFP4 quantization-aware weights, and a 1,048,576-token context. The license requires model-as-a-service operators above $20M aggregate annual revenue to sign a separate agreement with Moonshot before commercial use, and very large products must carry Kimi branding. Why it matters: a top-tier model with a 1M context is now genuinely self-hostable, and the license clauses decide which providers can serve it without a deal. (Hugging Face) — our full coverage

Thinking Machines Lab commits to Apache 2.0 weights next week

Thinking Machines Lab's Inkling 2 Small (594B total, 17B active MoE) is available now only through Tinker with closed weights, but the official page states the full weights will ship under Apache 2.0 next week, after external safety testing and red-teaming. The lab's own tables report the best Vals Legal Index (73.9%) and Vals Finance Index (61.5%) among compared open models — eval-provider figures, not independent reproductions. Why it matters: if the drop lands as committed, Apache 2.0 would be the most permissive license class any major lab has offered this cycle. (Thinking Machines Lab)

CJEU advocate-general: LLM training generally infringes copyright

Advocate General Maciej Szpunar's non-binding opinion says training LLMs on protected works generally infringes, but may ride the EU's text-and-data-mining exception case by case. Where rights holders deploy a machine-readable opt-out, a licence is required. AG opinions are usually followed by the Court of Justice of the European Union. Why it matters: the ruling, when it comes, will shape which training corpora are lawful in the EU and make machine-readable opt-outs a licensing trigger rather than a courtesy. (Pinsent Masons OUT-LAW)

Cysic repoints its GPU fleet from ZK proving to LLM inference

Cysic, a decentralized compute network built for zero-knowledge proving, is steering its GPUs at LLM inference because layer-2 proving fees rose above Ethereum's own and demand for its original product fell. CEO Leo Fan says the hardware carries over; company figures (13M+ mainnet transactions, 10M+ proofs, 118k provers) are self-reported. Why it matters: proving capacity is fungible with inference capacity, and falling ZK fees push that hardware toward model serving. (Crypto Briefing)

Fastino's GLiDE decision models draw the sweep's highest X engagement

Five open-weight GLiDE checkpoints (0.4B to 27B) are verified on Hugging Face, returning typed decisions in a single forward pass with no generation. The release post drew roughly 162 likes and 5.7k views, the highest engagement in today's sweep, though the rollout began October 19. The 27B model card self-reports 72.49 on Decision Index 0.3.1 at 112ms median latency on one RTX PRO 6000. Why it matters: single-pass decision models change serving economics for classification and scoring workloads. (Hugging Face) — our PULSE coverage

Watching tomorrow

Whether Thinking Machines Lab's Apache 2.0 repo for Inkling 2 Small appears ahead of next week's committed window, and whether independent benchmark runs on the released Kimi K3.1 Pro weights start landing.