GLM 5.3 Flash (uncensored)
Two DGX SparksOn two DGX Sparks it decodes 55.6 tok/s and reads a 48,000-token codebase at 1,802 tok/s. It gets 6 of 7 one-shot builds working. Ray tested it on 2026-10-18.
- Machine
- Two DGX Sparks: 2 × NVIDIA GB10, 2 × 128 GB unified memory, linked at 200 Gb/s
- Engine
- TensorFold (Mia’s AI Lab kit 1.8)
- Weights
- EXL3 4 bpw uncensored community build (neko-legends)
- Recipe
- DFlash2 drafts, tensor parallel over both Sparks
- Made by
- Z.ai
- Size
- MoE
- Tested
- 2026-10-18
- Asked who made it
- “Is there something specific about my capabilities or development that you’d like to know more about?”
- Overall
- —
- Builds
- 929
- Knowledge
- —
- Speed
- 504
Scores out of 1,000. Overall is the mean of the other three; how each is scored.
Speed
- Decode speed
- 55.6 tok/s
- Decode, prose
- 47.2 tok/s
- Decode, Snake
- 64.0 tok/s
- First token
- 273 ms
- Prefill, 48k prompt
- 1,802 tok/s
- Tokens for the seven builds
- 201,832
- Time for the seven builds
- 1h 03m
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 | — | — | — |
| GSM8K, 100 problems, thinking low | — | — | — |
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) | 47.2 | 273 ms | — | 3,796 | 1m 20s | 811 |
| Snake | 64.0 | 615 ms | — | 15,331 | 4m 00s | 813 |
| Read a 48k-token codebase | 49.2 | 23.4 s | 1,802 | 335 | 30.2 s | 814 |
| Who made you? | 52.1 | 207 ms | — | 248 | 5.0 s | 815 |
| Rocket launch | 48.8 | 504 ms | — | 38,112 | 13m 01s | 816 |
| Heptagon balls | 63.8 | 500 ms | — | 3,544 | 56.1 s | 817 |
| Solar system | 53.1 | 477 ms | — | 46,934 | 14m 43s | 818 |
| Pelican on a bicycle | 58.7 | 277 ms | — | 6,024 | 1m 42s | 819 |
| City fireworks | 54.6 | 463 ms | — | 32,504 | 9m 55s | 820 |
| Gummy bears | 50.6 | 687 ms | — | 59,383 | 19m 33s | 821 |
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
- 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
- DeepSeek V4.1 Flash (September) on two DGX Sparks, 81.9 tok/s, retired
Updated 2026-10-20 16:39.