Can I run Mixtral 8x7B (MoE) on NVIDIA RTX 5090 32GB?
Yes: Mixtral 8x7B (MoE) runs on NVIDIA RTX 5090 32GB with IQ4_XS at roughly 147 tokens/s. Estimated memory 25.7 GiB at 8,192 tokens of context.
With IQ4_XS you can push the context to about 17,664 tokens before memory runs out (F16 KV cache).
Memory by quantisation (8,192-token context)
| Quantisation | Memory | Verdict | Est. speed |
|---|---|---|---|
| F16Unquantized reference | 89.6 GiB | Does not fit | — |
| Q8_0Near-lossless | 48.8 GiB | Needs CPU offload | — |
| Q6_KVery close to F16 | 38.3 GiB | Needs CPU offload | — |
| Q5_K_MGreat quality/size | 33.5 GiB | Needs CPU offload | — |
| Q4_K_MBest default for most | 29.0 GiB | Tight | ~131 tok/s |
| IQ4_XSSmaller 4-bit (imatrix) | 25.7 GiB | Fits | ~147 tok/s |
| Q3_K_MNoticeable loss | 23.9 GiB | Fits | ~158 tok/s |
| Q2_KLast resort; big loss | 20.8 GiB | Fits | ~180 tok/s |
NVIDIA RTX 5090 32GB offers about 32 GiB of memory (94% usable for a model) at 1792 GB/s. Speed is a bandwidth-bound estimate of single-stream generation at 65% efficiency; real results vary by runtime, drivers and settings, so measure on your machine.
How this is calculated
Weights ≈ parameters × bits per weight ÷ 8. KV cache ≈ 2 × 32 layers × 8 KV heads × 128 head dim × 2 bytes per token. Decode speed ≈ efficiency × memory bandwidth ÷ bytes read per token (this is a mixture-of-experts model, so only the active parameters are read, but all of them must fit in memory). Learn the maths in VRAM math and memory bandwidth, or try the VRAM calculator.
Mixtral 8x7B (MoE) on other hardware
Similar models on NVIDIA RTX 5090 32GB
Running slowly or out of memory? See CUDA out of memory and the quantisation lesson.