4× RTX PRO 6000
235tok/sEveryday
1 request, prose prompt, context 19, with DSpark (per recipe) Source
Intelligence34with thinking
- Custom kernel required
Engine
- Engine
- vLLM 0.26.1rc1.dev303+g74295e3bd
- Quantization
- NVFP4 (Experten; Marlin W4A16), fp8 KV (block 256)
- Model family
- DeepSeek-V4-Flash
- Context
- 262,144
- Creator
- Infatoshi
- GitHub stars
- 15
- Repo updated
- Aug 25, 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
- 2 × NVIDIA RTX PRO 6000 Blackwell
- Weights
- Main weightsMJPansa/DeepSeek-V4-Flash-0731-NVFP4 MIT
- Engine
- vLLM 0.26.1rc1.dev303+g74295e3bd
- Context
- 262,144 tokens
Notes
What matters before you rebuild it.
- Custom kernel required
Sources
- github.com/Infatoshi/dsv4-flash-2x-rtxpro6000s/blob/main/docs/BENCHMARKS.md#concurrency-x-input-length-x-spec-vs-base
- github.com/Infatoshi/dsv4-flash-2x-rtxpro6000s/blob/main/docs/BENCHMARKS.md#long-context
- github.com/Infatoshi/dsv4-flash-2x-rtxpro6000s
- github.com/Infatoshi/dsv4-flash-2x-rtxpro6000s/blob/main/docs/BENCHMARKS.md
- github.com/Infatoshi/dsv4-flash-2x-rtxpro6000s/blob/main/scripts/serve_256k_marlin.sh
- github.com/Infatoshi/dsv4-flash-2x-rtxpro6000s/blob/main/bench/bench.py
Related recipes
2× DGX Spark
33.2tok/sEveryday
1 request, prose prompt, with DSpark (per recipe) Source
Intelligenceno independent value
- Custom kernel required
1× DGX Spark
30.2tok/sEveryday
1 request, story prompt, prompt 18 tokens, with MTP (per recipe) Source
Intelligence34with thinking
- ≤ 3 bit: quantization may cost quality
2× RTX 3090
21.9tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence34with thinking
- Partly in system RAM
- Experimental
1× DGX Spark
19.5tok/sEveryday
1 request, prose prompt, context 4096, no speculative decoding Source
Intelligence34with thinkingOriginal model's value. Heavily compressed (~2.1 bit), intelligence is likely much lower.
- ≤ 3 bit: quantization may cost quality
4× RTX PRO 6000
215tok/sMixed
1 request, mixed prompt set, with MTP Source
Intelligence24with thinkingno thinking 19
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.