Skip to content

Kimi K2 Instruct

Moonshot AI
Code Tool Calls

Kimi K2 Instruct is a 1,026.47-billion-parameter Mixture-of-Experts model from Moonshot AI, fine-tuned for instruction following, code generation, and autonomous tool use. It activates 8 of 384 experts plus 1 shared expert per token, achieving strong performance on coding and agentic benchmarks while keeping per-token compute equivalent to a 32B dense model. The model supports code generation and tool calling in English and Chinese, trained with the MuonClip optimizer on 15.5 trillion tokens. With a 128K context window and flash attention, it is suited for agentic deployments requiring reliable tool orchestration.

Hardware Configuration

Optional — for precise deployment recommendations
Quantization Quality Size Fit
Q8_0 High 1016.15 GB
Q8_K_XL High 1108.35 GB
Q6_K High 784.82 GB
Q6_K_XL High 818.73 GB
Q5_K_M Medium 678.37 GB
Q5_K_S Medium 658.07 GB
Q5_K_XL Medium 680.38 GB
Q4_K_M Medium 578.14 GB
Q4_K_S Medium 542.74 GB
Q4_K_XL Medium 546.77 GB
Q4_0 Medium 540.76 GB
Q4_1 Medium 598.41 GB
Q3_K_M Low 455.73 GB
Q3_K_S Low 412.01 GB
Q3_K_XL Low 421.03 GB
Q2_K Low 347.56 GB
Q2_K_L Low 347.82 GB
Q2_K_XL Low 355.65 GB
Last updated: March 5, 2026