Kimi K3.1 weights released: Gated MLA layers swapped for Gated DSA for long context
teortaxesTex · x · 2026-11-04
Kimi K3.1 weights are out. Per the technical notes, Kimi K3.1 replaces the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling. teortaxesTex had previously predicted Moonshot would make this MLA→DSA switch (with IndexCache) to build a fast everyday model competitive with DeepSeek v4.1 Flash pricing, and calls it another win for the domestic open ecosystem. Eval results default to reasoning effort max, temperature 1.0, top-p 0.95.
Related event: Moonshot AI's Kimi K3.1-Pro Hits Hugging Face with Open Weights(9 posts)→
More from Models
- Tencent reportedly rolls out AI model trained with limited cybersecurity data amid hacking fears — pstAsiatech · 2026-11-05
- Hydrogen-1.0, a science-tuned GLM-5.3 finetune, trends on Hugging Face — deep-principle · 2026-11-05
- Claude Opus wins sailing sim by writing autopilot code instead of steering — teortaxesTex · 2026-11-05
- AI semiotics: models are swayed by visual symbols without recognizing it in their chains-of-thought — emollick · 2026-11-05
- Liquid AI: vision encoders can burn more compute than the LLM on large images and video — JosephJacks_ · 2026-11-05
- Baseten explains how to make flash models faster: kernel fusion, PDL, and host loop cuts — baseten · 2026-11-05