Qwen 3.8 Flash Next
One DGX SparkOn one DGX Spark it decodes 66.7 tok/s and reads a 48,000-token codebase at 2,065 tok/s. It answers 80.7% of MMLU-Pro and gets 2 of 7 one-shot builds working. Ray tested it on 2026-10-19.
- Machine
- One DGX Spark: NVIDIA GB10, 128 GB unified memory
- Engine
- TensorFold
- Weights
- MLX 4-bit of the official weights (Vontra)
- Recipe
- MiaAI-Lab recipe, MTP 6 drafts, 220K context
- Made by
- Alibaba Qwen
- Size
- MoE
- Tested
- 2026-10-19
- Asked who made it
- “I am Qwen (Tongyi Qianwen), a large language model independently developed by Alibaba Group’s Tongyi Lab.”
- Overall
- 550
- Builds
- 286
- Knowledge
- 807
- Speed
- 558
Scores out of 1,000. Overall is the mean of the other three; how each is scored.
Speed
- Decode speed
- 66.7 tok/s
- Decode, prose
- 66.8 tok/s
- Decode, Snake
- 66.5 tok/s
- First token
- 282 ms
- Prefill, 48k prompt
- 2,065 tok/s
- Tokens for the seven builds
- 131,829
- Time for the seven builds
- 35m 54s
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% | 18.8 |
| GSM8K, 100 problems, thinking off | 99 of 100 | 99.0% | 3.6 |
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) | 66.8 | 282 ms | — | 2,128 | 32.1 s | 851 |
| Snake | 66.5 | 360 ms | — | 32,768 | 8m 12s | 853 |
| Read a 48k-token codebase | 59.8 | 22.2 s | 2,065 | 923 | 37.6 s | 854 |
| Who made you? | 78.4 | 248 ms | — | 84 | 1.3 s | 855 |
| Rocket launch | 61.0 | 291 ms | — | 15,273 | 4m 10s | 856 |
| Heptagon balls | 64.8 | 304 ms | — | 7,927 | 2m 02s | 857 |
| Solar system | 75.6 | 302 ms | — | 7,140 | 1m 34s | 858 |
| Pelican on a bicycle | 52.0 | 273 ms | — | 25,992 | 8m 19s | 862 |
| City fireworks | 67.9 | 289 ms | — | 8,254 | 2m 01s | 863 |
| Gummy bears | 60.4 | 440 ms | — | 34,475 | 9m 31s | 870 |
Other runs of Qwen 3.8 Flash Next
| Machine | Decode tok/s | Prose | Snake | Engine and weights |
|---|---|---|---|---|
| Two RTX 5090s | 218.5 | 164.8 | 272.2 | Strata, IQ2_XS |
| Mac Studio M3 Ultra | 100.9 | 78.8 | 123.0 | MLX 4.7 bpw |
| Two RTX 5090s | 69.6 | 69.6 | — | llama.cpp, IQ4_XS |
| Two DGX Sparks | 56.8 | 46.9 | 66.8 | vLLM NVFP4, 47K draft vocabulary |
More models on the bench
- Qwen 3.8 27B on one RTX 5090, 189.2 tok/s
- Qwen 3.8 27B on one DGX Spark, 31.8 tok/s
- Nemotron 3.5 Super on one DGX Spark, 46.0 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.