Run Codestral 2508 locally
Codestral 2508 is 22B parameters (dense). 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) | 13.2 GB | RTX 4090 (24 GB) ($0.34/hr) | smallest footprint, minor quality loss |
| int8 (Q8) | 26.4 GB | A100 (80 GB) ($1.19/hr) | near-lossless |
| fp16 (full) | 52.8 GB | A100 (80 GB) ($1.19/hr) | reference quality |
VRAM ≈ params × bytes/param × 1.2 (overhead). See the fullmethod and break-even calculator.
Self-host vs API
Cheapest API for Codestral 2508 is $1.2 / 1M tokens(blended) via Mistral AI. Running it yourself in int4 fits aRTX 4090 (24 GB) at $0.34/hr — about $248.2/month at 24/7. Those two lines cross at roughly 207M 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
No cloud account needed: Codestral 2508 in int4 (13.2 GB) fits a RTX 5090 (32 GB) + host PC (plus a host PC), a $3,500 one-off buy. Amortized over 3 years that's about $0.1332/hr — hardware only, electricity aside. Against the Mistral AI API at $1.2/1M, buying it pays for itself after roughly 2.9B tokens total. Below that the API is cheaper; a machine that mostly sits idle rarely earns back its price.
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