Kimi K3.1 Pro: Moonshot opens a 2.8T-parameter multimodal agent model
Moonshot AI published open weights for Kimi K3.1 Pro, a Mixture-of-Experts model with 2.8T total and 104B active parameters, a roughly 1M-token context window, and text-plus-image input, under a custom Kimi license. Its self-reported scores include 94.2 on GPQA-Diamond, 92.1 on BrowseComp and 79.6 on OSWorld V2.1, set against Claude Fable 5.1 and GPT-6 Astra. These are vendor figures with no independent verification yet, and at this size it is out of reach for local hardware, so it matters mostly as an open-weight option for hosted inference.
Sources & depth
Ripple-65B Preview: a small dense open-weight model claims the US open-weight coding lead
Ricursive Intelligence released Ripple-65B Preview, a dense 65B Transformer (80 layers, 12T training tokens, up to 1M context) derived from the model behind its chip-design platform. The lab reports 80.52 on TerminalBench 2.1, ahead of much larger open models such as Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4), and puts its full R&D cost at $32M. It is on Hugging Face, Together AI, GMI Cloud and OpenRouter. The scores are self-reported, and a dense 65B model is far easier to run than the trillion-parameter MoEs it is compared against.
Sources & depth
Tencent Hy4 appears on OpenRouter
Tencent's Hy4 model surfaced on Hacker News with a research page and on OpenRouter, priced at $0.417 per million input tokens and $1.251 per million output tokens through Tencent Cloud. The text-only API listing gives no context length, and Tencent's page offered no readable specifications, so size, architecture and benchmarks are unconfirmed.
Sources & depth
Anthropic shows a prototype for cryptographically proving which model served a response
Anthropic released FlashProve, an open-source research prototype that uses zero-knowledge proofs so a user can verify a response came from a specific committed model without seeing its weights. It generates proofs for Llama 4 Maverick (400B MoE) at about 110 tokens per second on one H100, but costs roughly 70 to 95 times the GPU time of serving the same model on vLLM and is limited to a 1,024-token context. The use case is checking that a provider is serving the model it claims and supporting compliance audits; it is not yet practical for production.
Sources & depth
Nvidia RTX Spark laptops run local inference several times slower on Linux
A MindStudio analysis finds the RTX Spark laptop is badly throttled on native Linux because power sharing depends on an encrypted firmware table that only the Windows driver can read. Plugged in, the GPU sits near 780 MHz versus about 2.2 GHz on Windows, and gpt-oss-120B runs at 13.9 tokens per second on Linux against 65.4 on Windows. WSL, which uses the Windows driver stack, matches Windows speed. A real fix needs Nvidia to add equivalent power management to its Linux driver, which matters to anyone buying the machine for local inference.
Sources & depth
Midterm results hit AI data center stocks
AI infrastructure stocks fell after Democrats won the Texas governor's race, where the winner backs an immediate moratorium on data centers. Wells Fargo estimates Texas holds about 27% of planned data center gigawatts, and analysts see political upside to blocking, delaying or raising the cost of new projects. SpaceX and Tesla fell about 4% and 5%, so local opposition to data centers is now a priced risk for AI compute buildout.
Sources & depth
OpenAI ships Images 2.6
OpenAI introduced Images 2.6, which it says brings faster image generation, GIF creation, more precise editing and new API controls across ChatGPT, Codex and the API. Details beyond OpenAI's own summary were not retrievable, so pricing and benchmark claims are unconfirmed.
Sources & depth