← CanIRunLocal · All hardware · All models
Yes, with partial offload (slower). Mixtral 8x7B (MoE) on RTX 4090 24GB at Q4_K_M uses system RAM (29.8 GB needed, 23.5 GB usable), about 10 tok/s (estimate).
Hardware: RTX 4090 24GB · Model: Mixtral 8x7B (MoE)
RTX 4090 24GB has 24 GB of VRAM and 1008 GB/s of memory bandwidth. The checker holds back 0.5 GB of VRAM for the driver and assumes 16 GB of system RAM, so fast memory usable by Ollama is 23.5 GB. Partial offload may also use 60% of that system RAM. Mixtral 8x7B (MoE) (46.7B, 12.9B active, Mistral) at Q4_K_M is a 28 GB download and needs 29.8 GB with headroom for about 8k tokens.
| Best quant | Q4_K_M |
|---|---|
| Download size | 28 GB |
| Memory needed | 29.8 GB |
| Usable fast memory | 23.5 GB |
| Fit | partial offload (slower) |
| Estimated speed | ~10 tok/s (estimate) |
| Command | ollama run mixtral:8x7b-instruct-v0.1-q4_K_M |
Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-23.
A project by OnDevice Labs · llms.txt