2× DGX Spark
36.1tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence42with thinking
- ≤ 3 bit: quantization may cost quality
- Custom kernel required
- Non-commercial
- License unclear
Engine
- Engine
- vLLM
- Quantization
- TrellisMX K4/K5 Experten → FP8 E4M3 Math (UE8M0/32) + nvfp4_ds_mla KV
- Model family
- GLM-5.3-Flash
- Context
- 1,000,000
- Creator
- brandonmusic
- GitHub stars
- 8
- Repo updated
- Sep 10, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
No sourced measurements for this recipe.
What you need
- Hardware
- 4 × NVIDIA RTX PRO 6000 Blackwell
- Weights
- Main weightsbrandonmusic/GLM-5.3-Flash-TrellisMX-MXFP8 other
- Basezai-org/GLM-5.3-Flash-BF16 MIT
- Engine
- vLLM
- Context
- 1,000,000 tokens
Notes
What matters before you rebuild it.
- Custom runtime
- License unclear
Sources
Related recipes
4× DGX Spark
34.1tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence42with thinking
- Custom kernel required
- Non-commercial
4× DGX Spark
31.5tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence42with thinking
- Custom kernel required
- Non-commercial
1× DGX Spark
19.5tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence42with thinkingOriginal model's value. Heavily compressed (~2.3 bit), intelligence is likely much lower.
- ≤ 3 bit: quantization may cost quality
- Experimental
- Custom kernel required
- Non-commercial
2× DGX Spark
19.1tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence42with thinking
- Non-commercial
- License unclear
2× RTX 3090
18.1tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence42with thinking
- ≤ 3 bit: quantization may cost quality
- Partly in system RAM
Local AI in your company?
In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.