Moonshot AI releases open Kimi K3, a 2.8T-parameter MoE model with 1M context

Moonshot AI has released Kimi K3, an open-source sparse Mixture-of-Experts (MoE) model featuring native vision capabilities and a 1-million-token context window. It is the first open model to reach the 2.8 trillion parameter scale.

• The architecture introduces Kimi Delta Attention (KDA) for up to 6.3x faster decoding in long contexts and Attention Residuals (AttnRes) for roughly 25% higher training efficiency.
• K3 activates only 16 of 896 experts, achieving roughly 2.5x better overall scaling efficiency than its predecessor, Kimi K2.
• In benchmarks, K3 leads Program Bench, SWE Marathon, BrowseComp, Automation Bench, and OmniDocBench, while trailing Claude Fable 5 on FrontierSWE and HLE-Full.
• The model is accessible via the OpenAI SDK with flat pricing of $0.30/MTok for cache-hit input and $15.00/MTok for output.

Moonshot states that K3 targets long-horizon coding, knowledge work, and reasoning, offering a high-performance open alternative to proprietary models like Claude Fable 5 and GPT 5.6 Sol.