Moonshot releases open-weight Kimi K3.1 Pro: 2.8T params, 104B active, 1M context
MaziyarPanahi · x · 2026-11-05
Moonshot AI published Kimi K3.1 Pro on Hugging Face, an open-weight, natively multimodal agentic model post-trained from Kimi K3 with joint RL across hundreds of downstream domains. It uses a MoE architecture with 2.8T total / 104B active parameters, 93 layers, and a 1M-token context window. Compared to K3, full-attention Gated MLA layers were partly replaced with Gated DeepSeek Sparse Attention (Gated DSA) for more efficient long-context modeling; it uses SiTU-GLU activation, a 401M-param MoonViT-V2 vision encoder, and MXFP4/MXFP8 quantization-aware training. The license permits fine-tuning and derivative works.
Related event: Moonshot AI Open-Sources Kimi K3.1 Pro: 2.8T-Param MoE with 1M Context(24 posts)→
More from Models
- Ripple claims $32M compute for open-weight coding SOTA; skeptics question the math — teortaxesTex · 2026-11-05
- Gemini 4 Flash (codename Hydrogen) spotted A/B testing in AI Studio and LMArena — lyraxana · 2026-11-05
- Perceptron's new 3D hand model beats ACE-Ego-Hand by 12-25% on HOT3D and ARCTIC — gui_penedo · 2026-11-05
- 0.8B open-source HaloGuard 2.0 claims streaming-guard SOTA, beating models 5-10x its size — const_reborn · 2026-11-05
- MolmoAct 2 Takes Second to GPT-6 Astra in RoboColosseum's Real-World Robot Tests — DJiafei · 2026-11-05
- Claude Max subscriber at $100/month says Opus 5.5 'almost feels unlimited' — prasenx · 2026-11-05