){kind=icon}
What is this?
Kimi K3 is MoonshotAI's open-weight, native multimodal agentic model — a 2.8T-parameter Mixture-of-Experts system built for frontier-level long-horizon coding, agentic knowledge work, and reasoning.
Kimi K3 is MoonshotAI's open-weight, native multimodal agentic model — a 2.8T-parameter Mixture-of-Experts system built for frontier-level long-horizon coding, agentic knowledge work, and reasoning.Is this a hosted chatbot I can just use, or something I have to run myself?It's a release of raw model weights, not a hosted product — you download and serve the weights yourself under the Kimi K3 License.
It's a release of raw model weights, not a hosted product — you download and serve the weights yourself under the Kimi K3 License.What hardware does running the full model actually require?Only 104B of the model's 2.8T total parameters are activated per token thanks to its sparse MoE design, and weights use MXFP4/MXFP8 quantization-aware training to shrink the footprint, but you still need multi-GPU infrastructure to serve a model this size.
Only 104B of the model's 2.8T total parameters are activated per token thanks to its sparse MoE design, and weights use MXFP4/MXFP8 quantization-aware training to shrink the footprint, but you still need multi-GPU infrastructure to serve a model this size.How sparse is the model really — does it use all its experts at once?No, it only activates 16 of 896 experts per token via the Stable LatentMoE framework, which is what keeps the effective compute manageable despite the 2.8T parameter count.
No, it only activates 16 of 896 experts per token via the Stable LatentMoE framework, which is what keeps the effective compute manageable despite the 2.8T parameter count.What is Kimi Delta Attention (KDA)?KDA is the new attention mechanism underlying K3, used in 69 of its 93 layers alongside Gated MLA in the remaining layers, paired with Attention Residuals (AttnRes) to improve scaling efficiency.
KDA is the new attention mechanism underlying K3, used in 69 of its 93 layers alongside Gated MLA in the remaining layers, paired with Attention Residuals (AttnRes) to improve scaling efficiency.How much better is K3 than its predecessor Kimi K2?MoonshotAI reports roughly a 2.5x improvement in overall scaling efficiency over Kimi K2, driven by the new Stable LatentMoE sparsity design.
MoonshotAI reports roughly a 2.5x improvement in overall scaling efficiency over Kimi K2, driven by the new Stable LatentMoE sparsity design.Can it understand images and video, or just text?K3 is natively multimodal: it processes text, images, and video within the same model using a 401M-parameter MoonViT-V2 vision encoder, plus a 1-million-token context window.
K3 is natively multimodal: it processes text, images, and video within the same model using a 401M-parameter MoonViT-V2 vision encoder, plus a 1-million-token context window.How does it compare to Claude and GPT on coding benchmarks?On Terminal-Bench 2.1 it scores 88.3, ahead of Claude Opus 4.8 (84.6) and GPT-5.5 (83.4), and it leads on SWE-Marathon (42.0) among the compared frontier models.
On Terminal-Bench 2.1 it scores 88.3, ahead of Claude Opus 4.8 (84.6) and GPT-5.5 (83.4), and it leads on SWE-Marathon (42.0) among the compared frontier models.What are similar projects?DeepSeek-V3, ZDTaichu5.0-9B, Qwen3.8-Flash-Next, StartLux-Decision, Kimi-K2, Qwen3.8, Qwen3-VL, jeeves
DeepSeek-V3, ZDTaichu5.0-9B, Qwen3.8-Flash-Next, StartLux-Decision, Kimi-K2, Qwen3.8, Qwen3-VL, jeevesWhich topics is it in?long-horizon-agents, deep-research, open-weights, model, foundation-models, vision-language-model, mixture-of-experts, reasoning, language-model, moe, quantization, multimodal
long-horizon-agents, deep-research, open-weights, model, foundation-models, vision-language-model, mixture-of-experts, reasoning, language-model, moe, quantization, multimodal