Qwen 3.8 27B
One RTX 5090On one RTX 5090 it decodes 189.2 tok/s and reads a 48,000-token codebase at 4,057 tok/s. It answers 80.7% of MMLU-Pro and gets 6 of 7 one-shot builds working. Ray tested it on 2026-10-20.
- Machine
- One RTX 5090: 32 GB GDDR7, one card of Ray’s dual-5090 desktop
- Engine
- NInfer
- Weights
- NVFP4, official weights
- Recipe
- DFlash2 drafts (7), FP8 KV cache, 131K context
- Made by
- Alibaba Qwen
- Size
- 27B dense
- Tested
- 2026-10-20
- Asked who made it
- “I was created by Alibaba Group’s Tongyi Lab.”
- Note
- First-token times here include about 3 s that Ownlight’s gateway adds before it passes a request to the dual-5090 desktop. Ownlight’s own traces for these runs show 0.13 to 0.43 s to the first token, and 5,535 tok/s prefill on the 48k prompt.
- Overall
- 843
- Builds
- 857
- Knowledge
- 807
- Speed
- 864
Scores out of 1,000. Overall is the mean of the other three; how each is scored.
Speed
- Decode speed
- 189.2 tok/s
- Decode, prose
- 169.1 tok/s
- Decode, Snake
- 209.2 tok/s
- First token
- 3.2 s
- Prefill, 48k prompt
- 4,057 tok/s
- Tokens for the seven builds
- 191,508
- Time for the seven builds
- 17m 50s
Decode speed is the median of a prose prompt and a Snake game, one request at a time. Prefill is how fast it reads a 48,000-token codebase before it answers.
Accuracy
| Set | Correct | Score | Minutes |
|---|---|---|---|
| MMLU-Pro, 140 questions, thinking low | 113 of 140 | 80.7% | 6.9 |
| GSM8K, 100 problems, thinking off | 97 of 100 | 97.0% | 3.5 |
Builds
Run these pagesRun local models in Ownlight
Ownlight connects the model servers you already run, like Ollama, LM Studio, Unsloth Studio, llama-swap or Model Deck, and starts Claude Code or Codex on a local model beside your cloud agents. Local is part of Free.
Every step
Decode speed, time to first token, prefill speed, output tokens and total time for each prompt in the suite.
| Test | Decode tok/s | First token | Prefill tok/s | Output tokens | Total time | Run |
|---|---|---|---|---|---|---|
| Prose (speed) | 169.1 | 3.2 s | — | 2,213 | 15.9 s | 917 |
| Snake | 209.2 | 3.3 s | — | 16,466 | 1m 21s | 919 |
| Read a 48k-token codebase | 205.2 | 11.3 s | 4,057 | 714 | 14.8 s | 920 |
| Who made you? | 333.3 | 3.2 s | — | 44 | 3.3 s | 921 |
| Rocket launch | 182.6 | 3.2 s | — | 59,921 | 5m 31s | 922 |
| Heptagon balls | 268.0 | 3.2 s | — | 6,139 | 26.1 s | 923 |
| Solar system | 263.3 | 3.1 s | — | 4,179 | 19.0 s | 924 |
| Pelican on a bicycle | 284.5 | 3.2 s | — | 11,059 | 42.1 s | 925 |
| City fireworks | 256.3 | 3.3 s | — | 5,507 | 24.8 s | 926 |
| Gummy bears | 162.9 | 3.4 s | — | 88,237 | 9m 04s | 927 |
| Capybara onsen | 156.3 | 3.3 s | — | 78,304 | 8m 24s | 928 |
Other runs of Qwen 3.8 27B
| Machine | Decode tok/s | Prose | Snake | Engine and weights |
|---|---|---|---|---|
| Mac Studio M3 Ultra | 47.0 | 47.0 | — | llama.cpp, Q4_0 + DFlash2 drafts |
| One DGX Spark | 31.8 | 25.7 | 37.9 | SGLang, NVFP4, official weights |
| One DGX Spark | 12.4 | 12.4 | 12.4 | SGLang NVFP4, no draft model |
More models on the bench
- Qwen 3.8 27B on one DGX Spark, 31.8 tok/s
- Nemotron 3.5 Super on one DGX Spark, 46.0 tok/s
- Qwen 3.8 Flash Next on one DGX Spark, 66.7 tok/s
- DeepSeek V4.1 Flash 1017 on two DGX Sparks, 78.6 tok/s
- GLM 5.3 Flash (uncensored) on two DGX Sparks, 55.6 tok/s
- DeepSeek V4.1 Flash (September) on two DGX Sparks, 81.9 tok/s, retired
Updated 2026-10-20 16:39.