Skip to content

Can RTX 4090 run Qwen3 8B Instruct?

Comfortable: GGUF Q6_K needs 7.6 GB at 4K, batch 1; headroom 16.4 GB.

Text input, FP16 KV cache, one card. Unified-memory capacity uses the GPU guide’s usable limit; negative headroom may require host offload. Speed source settings may differ from this 4K estimate.

Measured runs

FrameworkQuanttok/sVRAM used (GB)
llama.cppGGUF Q4_K_M1425.8

Quant memory at 4K

QuantTotal GBVerdictWeight sizing
GGUF Q4_K_M5.8ComfortableEstimated weights
GGUF Q6_K7.6ComfortableEstimated weights
AWQ INT44.9ComfortableEstimated weights
EXL2 4.65bpw5.6ComfortableEstimated weights

Context capacity · GGUF Q6_K

ContextTotal GBVerdict
4K7.6Comfortable
16K9.5Comfortable
32K12.0Comfortable

Generation speed

Measured: 142 tok/s, GGUF Q4_K_M, llama.cpp

GGUF Q6_K: 105–139 tok/s · Derived range

Other model / GPU pairs