back

Kimi K2

Kimi

Moonshot AI · 1T (32B active) · Mixture of Experts

1T-param MoE with 384 experts — 32B active, strong agentic coding

110.1K downloads 2.3K likes 2025-07 128K context

Use Cases

chat reasoning code

Mixture of Experts

Total experts: 384
Active experts: 8
Active params: 32.0B

Quantization Options

Quant Bits VRAM Quality Status
Q2_K 2 320.6 GB low
Q3_K_M 3 448.7 GB moderate
Q4_K_M 4 512.7 GB good
Q5_K_M 5 640.8 GB good
Q6_K 6 768.8 GB excellent
Q8_0 8 1025 GB excellent
F16 16 2049.4 GB lossless

About this model

Kimi K2-Instruct-0905 is the latest, most capable version of Kimi K2. It is a state-of-the-art mixture-of-experts (MoE) language model, featuring 32 billion activated parameters and a total of 1 trillion parameters.

Key Features

  • Enhanced agentic coding intelligence: Kimi K2-Instruct-0905 demonstrates significant improvements in performance on public benchmarks and real-world coding agent tasks.
  • Improved frontend coding experience: Kimi K2-Instruct-0905 offers advancements in both the aesthetics and practicality of frontend programming.
  • Extended context length: Kimi K2-Instruct-0905’s context window has been increased from 128k to 256k tokens, providing better support for long-horizon tasks.