Kimi Linear 48B A3B Instruct
Moonshot AI · 2025-10-30 · 49.1B parameters
Kimi Linear 48B A3B Instruct is the instruction-tuned variant of Moonshot AI's Kimi Linear hybrid linear-attention MoE model with 48B total and 3B active parameters, released October 2025. Fine-tuned for chat and instruction-following on top of the 5.7T-token linear-attention base, it inherits the architecture's key benefits of fast decoding, reduced KV cache usage, and strong long-context performance.