Moonshot AI published Kimi K3.1 Pro weights — an open-weight, natively multimodal Mixture-of-Experts agentic model (2.8T params, 104B activated, 1M-token context) with a MoonViT vision encoder and quantization-aware weights, claiming strong gains on coding, agentic, multimodal, and long-context benchmarks. The release highlights architecture details (896 experts, Gated DSA for long context), quantization (MXFP4/MXFP8), and extensive comparative evaluation against leading models.

The post summarizes the September changes merged into Servo 0.7.0 (527 commits), highlighting many new web platform features (e.g., <progress>, :checked, user-select, CSS typed arithmetic, roundRect), numerous DOM API additions, and upgrades to SpiderMonkey and mozangle for security. It also covers embedding API changes, accessibility and media/text-track improvements, WebGL2/WebCrypto work, and a raft of performance and memory optimizations.