Qwen3.8 27B on the NVIDIA RTX 5090
Qwen3.8 27B runs at about 76.9 tokens/s on the NVIDIA RTX 5090 (measured, Q4_K_M, CUDA, single-stream llama-server decode).
Best measured: 76.9 tokens/s generation (Q4_K_M, CUDA). The source gives no prompt-processing figure.
For comparison, our size-scaling estimate for this combination is 67.1 tok/s.
Every run
| Quant | Backend | Generate tok/s | Prompt tok/s | Test | Build / by / notes | Source |
|---|---|---|---|---|---|---|
| Q4_K_M | CUDA | 76.9 | – | 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) |
| UD-Q4_K_XL | CUDA | 71.9 | – | 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 | 71.9 | – | 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 | 71.9 | – | 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 | 71.9 | – | 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 | 71.9 | – | 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 | 71.9 | – | 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 | 71.9 | – | 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-Q5_K_XL | CUDA | 69.6 | – | 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-Q5_K_XL | CUDA | 69.6 | – | 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) |
| Q6_K | CUDA | 62.0 | – | 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. 256K context. | qwen38-mtp community table (GitHub) |
| Q6_K | CUDA | 62.0 | – | 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. 256K context. | qwen38-mtp community table (GitHub) |
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).
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 5090 · Qwen3.8 27B on other hardware
Embed this result
HTML
<a href="https://rigbench.pages.dev/model/qwen3-8-27b/on/nvidia-rtx-5090/"><img src="https://rigbench.pages.dev/badge/qwen3-8-27b--nvidia-rtx-5090.svg" alt="Qwen3.8 27B · NVIDIA RTX 5090: 76.9 tok/s (Q4_K_M, measured)"></a>
Markdown (GitHub README, forums)
[](https://rigbench.pages.dev/model/qwen3-8-27b/on/nvidia-rtx-5090/)