Run Kimi K2 Thinking locally
Kimi K2 Thinking is 1000B parameters (32B active per token — a mixture-of-experts, so it still needs room for all 1000B in memory). Here's the VRAM it needs at each quantization, the smallest GPU that fits, and how a self-hosted box compares with the cheapest verified API.
VRAM by quantization
| Quantization | VRAM needed | Fits on | Notes |
|---|---|---|---|
| int4 (Q4) | 600 GB | 8× H100 (640 GB) ($15.92/hr) | smallest footprint, minor quality loss |
| int8 (Q8) | 1200 GB | multi-node (>640 GB) | near-lossless |
| fp16 (full) | 2400 GB | multi-node (>640 GB) | reference quality |
VRAM ≈ params × bytes/param × 1.2 (overhead). See the fullmethod and break-even calculator.
Self-host vs API
Cheapest API for Kimi K2 Thinking is $3.1 / 1M tokens(blended) via Google. Running it yourself in int4 fits a8× H100 (640 GB) at $15.92/hr — about $11,621.6/month at 24/7. Those two lines cross at roughly 3.7B tokens/month: below that the API wins on cost, above it the dedicated GPU does (assuming you keep it busy). Tune your own volume in the break-even calculator.
Run it at home
In int4, Kimi K2 Thinking needs 600 GB — more than any consumer desktop or Apple machine we track holds, so at home it means a multi-GPU server. Renting cloud GPUs by the hour, or the API, is the practical route. See the machines table for the ceiling.
Get the weights
Quantized builds on the model hubs — links search live, so they track new builds as they appear:
- GGUF builds llama.cpp · Ollama · LM Studio
- AWQ builds GPU serving (vLLM)
- GPTQ builds GPU serving
- MLX builds Apple silicon
- Ollama library one-command local run
License: open weights under Modified MIT · terms. Compare all providers on the kimi-k2-thinking API page.