DeAI Daily Brief — 24 October 2026

Today in DeAI: Moonshot AI's Kimi K3.1 Pro hits the API at $2/$10 per million tokens with open weights promised by November 4, You.com routes 12 open-weight models behind an enterprise SLA, and Mubit Labs ships a 23B verification model under Apache 2.0.

Moonshot ships Kimi K3.1 Pro on the API; open weights due November 4

Moonshot AI's Kimi K3.1 Pro, a post-trained version of the 2.8-trillion-parameter Kimi K3 mixture-of-experts model, is live on the Kimi API at $2.00/$10.00 per million input/output tokens, and the company states full model weights will be released by November 4, 2026. The model scores 55 on Artificial Analysis' independent Intelligence Index, ahead of GPT-6 Astra and Gemini 4 Argon at 53. Why it matters: if the weights ship as promised, the highest-scoring open-weight model available would sit roughly three points behind the strongest closed model on the index. (Moonshot AI) — our full coverage

You.com routes 12 open-weight models behind an enterprise SLA

You.com's Inference Router, launched October 22, exposes 12 open-weight models through one OpenAI-compatible endpoint with US hosting, and the company claims a 99.9% uptime SLA and zero data retention. Both commitments are provider claims without published audits so far. Why it matters: SLA-backed open-weight routing attacks the reliability gap that keeps compliance-sensitive workloads on closed hyperscaler APIs. (Business Wire) — our PULSE coverage

Mubit's Banyan Sapling is a 23B verifier that says unsure 45% of the time

Mubit Labs released Banyan Sapling on Hugging Face under Apache 2.0: a 23B model pruned from Qwen3-30B-A3B that reads a policy, evidence, and a claim and returns a calibrated verdict. Its model card reports all 779 calibrated answers correct, with 952 of 1,731 evaluation cases returned as unsure. Why it matters: the open-weights wave is specializing into small auditable components for agent trust pipelines, and this one is sized for a single 48GB-class GPU. (Hugging Face) — our PULSE coverage

Google Cloud makes Confidential Computing GA on Blackwell G4 instances

Google Cloud moved Confidential VMs and Confidential GKE Nodes to general availability on G4 instances with NVIDIA RTX PRO 6000 Blackwell GPUs, per the company's blog, putting hardware-backed TEE protection on current-generation GPU instances without code changes. Why it matters: attested, memory-encrypted inference on mainstream accelerators turns "verify what the provider can see" from an architecture project into a checkbox, and raises the private-execution bar every provider gets compared against. (Google Cloud) — see our private-inference explainer

NVIDIA's Nemotron 3.5 Bolt targets 24GB local GPUs

NVIDIA's Nemotron 3.5 pair splits by hardware: Super VL (120B total, 12B active, 1M context) scores 26 on Artificial Analysis' index, while Bolt 74B-A9B, a hybrid Mamba-Transformer, is engineered for agentic workloads on a 24GB RTX or DG Spark with native 2-bit quantization, per NVIDIA's claims. Why it matters: Bolt concedes most self-hosters cannot load 500B-class MoEs and optimizes for hardware people actually own. (NVIDIA)

Watching tomorrow

Mistral Large 4's open weights are committed for October 27, three days out; the license text and independent benchmark scores land first.