How fast does Qwen3.8 27B run on the NVIDIA RTX 4090?

Qwen3.8 27B runs at about 47.9 tokens/s on the NVIDIA RTX 4090 (measured, UD-Q4_K_XL, CUDA, llama-bench). Based on 7 measured run(s), each linked to its public source.

Does Qwen3.8 27B UD-Q4_K_XL fit in the NVIDIA RTX 4090's memory?

Estimate: the UD-Q4_K_XL file is 17.6 GB and needs about 19.0 GB with an ~8K context; the NVIDIA RTX 4090 has about 23.5 GB usable, so it fits.

Qwen3.8 27B › NVIDIA RTX 4090

Qwen3.8 27B on the NVIDIA RTX 4090

Qwen3.8 27B runs at about 47.9 tokens/s on the NVIDIA RTX 4090 (measured, UD-Q4_K_XL, CUDA, llama-bench).

Best measured: 47.9 tokens/s generation (UD-Q4_K_XL, CUDA). The source gives no prompt-processing figure.

Every run

Quant	Backend	Generate tok/s	Prompt tok/s	Test	Build / by / notes	Source
UD-Q4_K_XL	CUDA	47.9	–	tg (llama-bench)	b11200 unsloth UD quants; generation length not stated.	ComputingForGeeks Qwen3.8-27B test
Q4_K_M		47.7	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. File size 17,106,775,008 bytes as stated by the source.	qwen38-mtp community table (GitHub)
Q4_K_M	CUDA	47.0	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench. 500 W power limit. File size 17,106,775,008 bytes as stated by the source.	qwen38-mtp community table (GitHub)
UD-Q5_K_XL	CUDA	41.5	–	tg (llama-bench)	b11200 unsloth UD quants; generation length not stated.	ComputingForGeeks Qwen3.8-27B test
UD-Q4_K_XL	CUDA	44.7	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench.	qwen38-mtp community table (GitHub)
UD-Q4_K_XL	CUDA	44.7	–	decode, llama-server, MTP off	Baseline (spec-off) column of a paired MTP A/B: llama-server single stream (--parallel 1), probe.py median, not llama-bench.	qwen38-mtp community table (GitHub)
UD-Q6_K	CUDA	39.6	–	tg (llama-bench)	b11200 unsloth UD quants; generation length not stated.	ComputingForGeeks Qwen3.8-27B test

tg128 = generating 128 tokens; pp512 = reading a 512-token prompt (llama-bench). LocalScore averages several prompt and output lengths, so its numbers run lower than tg128. "decode, llama-server" = single-stream generation speed measured against a running llama-server (no speculative decoding).

Check prices

Amazon Newegg

As an Amazon Associate I earn from qualifying purchases.

No card yet?

Try it on a rented GPU (RunPod)

Affiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.

All results for the NVIDIA RTX 4090 · Qwen3.8 27B on other hardware

Embed this result

HTML

<a href="https://rigbench.pages.dev/model/qwen3-8-27b/on/nvidia-rtx-4090/"><img src="https://rigbench.pages.dev/badge/qwen3-8-27b--nvidia-rtx-4090.svg" alt="Qwen3.8 27B · NVIDIA RTX 4090: 47.9 tok/s (UD-Q4_K_XL, measured)"></a>

Markdown (GitHub README, forums)

[![Qwen3.8 27B · NVIDIA RTX 4090: 47.9 tok/s (UD-Q4_K_XL, measured)](https://rigbench.pages.dev/badge/qwen3-8-27b--nvidia-rtx-4090.svg)](https://rigbench.pages.dev/model/qwen3-8-27b/on/nvidia-rtx-4090/)