Moonshot AI released Kimi K3.1 Pro as an open-weight, natively multimodal agentic model with 2.8T total parameters, 104B activated parameters, and a 1M-token context window. Built from Kimi K3 using large-scale joint reinforcement learning, it supports text and images and uses MoE, KDA, and Gated DSA architecture with MXFP4/MXFP8 quantization. It is available through OpenAI/Anthropic-compatible APIs and vLLM, SGLang, and TokenSpeed, achieving 65.7 on Terminal-Bench 4.0 and 94.2 on GPQA-Diamond.

Images 2.6 is available in ChatGPT, Codex, and the API, generating images 50% faster and supporting GIF animation. API users can choose GPT‑Image‑2.6 Flare for fast generation at low quality or Sunburst for higher-quality, more consistent edits. The update adds image markup to the API, SVG exports, and a complexity parameter separate from image quality to control planning effort. It is globally available across ChatGPT and Codex subscription tiers, with both API models offered to developers.

Ricursive Intelligence released Ripple-65B Preview, a 65B dense Transformer trained on 12 trillion tokens with up to 1M context, claiming a U.S. open-weight SOTA TerminalBench 2.1 score of 80.52 for agentic coding. It surpasses larger listed models and runs efficiently on local NVIDIA hardware. Ricursive says full R&D and training cost $32M, used 1.4M RL rollouts, and achieved comparable code performance to the Olmo 3 ladder with 16x less compute through architecture, data, SFT, and RL optimization.

Spec Kit is an open-source toolkit that gives AI coding agents structured, reusable processes for building features, fixing bugs, and assessing ideas. Its core Spec-Driven Development flow turns specifications into plans, tasks, code, and convergence checks, while installable extensions support bug diagnosis and evidence-based decisions. Integrations, presets, workflows, bundles, catalogs, and project-local overrides let teams customize or create processes, with documented artifacts and support for offline or private environments.

Anthropic and collaborators released FlashProve, an open-source prototype for provable inference using zkSNARKs. It commits to a model’s hidden weights and proves responses came from that committed model, allowing CPU verification without rerunning inference. On a single H100, it proves Llama 4 Maverick (400B-parameter MoE) at about 110 tokens/second, with 19 MB proofs per 1,024-token request—roughly 70–95× vLLM’s GPU time. The release includes C++ and Lean verifiers, formalized security claims, code, and documentation; it hides weights but not architecture and supports integer greedy decoding.