Moonshot AI · 1T (32B active) · Mixture of Experts
1T-param MoE with 384 experts — 32B active, strong agentic coding
161.8K downloads
2.4K likes
2025-07 128K context
Use Cases
chat reasoning code
Mixture of Experts
Total experts: 384
Active experts: 8
Active params: 32.0B
| 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 | — |
GPUs that run Kimi K2
Kimi K2 needs about 512.7 GB of VRAM at Q4_K_M — more than any single GPU we track. Running it takes multiple GPUs, a large unified-memory machine, or CPU offload with plenty of system RAM.
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.