2× RTX 5090
149tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligenceno independent value
- Experimental
- License unclear
Local AI · Recipe · 4× RTX 3090
Nemotron-3-Puzzle 75.3B NVFP4 with vLLM on 4× RTX 3090
Nemotron-3-Puzzle 75.3B NVFP4 with vLLM on 4× RTX 3090: 104 tok/s according to github.com (dataset as of Nov 3, 2026).
by noonghunna
Intelligence (original model) · Artificial Analysisno independent value
Engine
- Engine
- vLLM
- Quantization
- NVIDIA modelopt NVFP4 (Experten) + FP8 (Mamba/Shared) → auf Ampere Marlin-W4A16-Dequant; fp8_e4m3 KV
- Model family
- Nemotron-3-Puzzle
- Context
- 200,000
- Parameters
- 75.3B MoE / 9.3B aktiv (Mamba2-Hybrid)
- Creator
- noonghunna
- GitHub stars
- 2,471
- Repo updated
- Nov 2, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
Decode · 1 request
What you need
- Hardware
- 4 × NVIDIA RTX 3090
- Weights
- Main weightsnvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 other
- Engine
- vLLM
- Context
- 200,000 tokens
Notes
What matters before you rebuild it.
- Experimental
- License unclear
Sources
Related recipes
2× RTX 3090
45.8tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligenceno independent value
- Experimental
- License unclear
4× RTX 3090
145tok/sEveryday
1 request, prose prompt, with DFlash (per recipe) Source
Intelligence34with thinkingno thinking 20
- Experimental
4× RTX 3090
110tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence40with thinking
- License unclear
4× RTX 3090
105tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence21with thinkingno thinking 20
4× RTX 3090
85.3tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence34with thinkingno thinking 20
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.