Can the GeForce RTX 3090 (used) run Mixtral 8x7B Instruct v0.1?
Only with CPU offload at Q4_K_M.
Mixtral 8x7B Instruct v0.1 is 46.7B parameters (MoE; all experts resident). At Q4_K_M the weights are about 26.1 GiB; with the 1.5 GiB runtime reserve that is 27.6 GiB against the card's 24GB, about 3.6 GiB more than the card has.
All three quants
24GB card| Quant | Weights | With reserve | Verdict | Headroom |
|---|---|---|---|---|
| Q4_K_Mthe everyday quant | 26.1 GiB | 27.6 GiB | Only with CPU offload | short 3.6 GiB |
| Q8_0near-lossless | 46.1 GiB | 47.6 GiB | Only with CPU offload | short 23.6 GiB |
| FP16full weights | 87.0 GiB | 88.5 GiB | Does not fit | short 64.5 GiB |
Verdict rules: fits comfortably when weights plus reserve sit within 85% of VRAM; tight when they fit with little left for context; CPU offload when they exceed VRAM by up to 2x (runs, slowly, with layers in system RAM); no beyond that. The KV cache per token depends on the model's architecture and is not modelled here.
Not this card. These fit it comfortably at Q4.
Nothing in the consumer catalog fits Mixtral 8x7B Instruct v0.1 at Q4 on a single card; this is multi-GPU or datacenter territory, or a job for hosted inference (see PicksByModel for per-token pricing).
What the GeForce RTX 3090 (used) does run well → · Mixtral 8x7B Instruct v0.1 on every card →
About the model
Mixtral 8x7B Instruct v0.1 by Mistral AI: 32,768-token native context, text, released 2023 under the Apache-2.0. Model card →