4× DGX Spark
87.7tok/sEveryday
1 request, prose prompt, with DSpark (per recipe) Source
Intelligence40with thinkingno thinking 25
Local AI · Recipe · 4× DGX Spark
DeepSeek-V4.1-Flash 763B EXL3 3.5 bpw with vLLM on 4× DGX Spark
DeepSeek-V4.1-Flash 763B EXL3 3.5 bpw with vLLM on 4× DGX Spark: 39.4 tok/s according to github.com (dataset as of Nov 3, 2026).
by tonyd2wild
Intelligence (original model) · Artificial Analysis40with thinkingno thinking 25
Engine
- Engine
- vLLM + cuda-exl3 (Image vllm-dsv41:exl3a, Patch-Set tp3e)
- Quantization
- EXL3 3.5bpw routed experts + FP8 Rest
- Model family
- DeepSeek-V4.1-Flash
- Context
- 300,000
- Parameters
- ~763B
- Creator
- tonyd2wild
- GitHub stars
- 154
- Repo updated
- Oct 26, 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
Prefill by context
What you need
- Hardware
- 4 × NVIDIA DGX Spark (GB10)
- Weights
- Main weightsbot-lab-21/DeepSeek-V4.1-Flash-EXL3-3.5bpw-Pollard MIT
- Basedeepseek-ai/DeepSeek-V4.1-Flash MIT
- Engine
- vLLM + cuda-exl3 (Image vllm-dsv41:exl3a, Patch-Set tp3e)
- Context
- 300,000 tokens
Notes
What matters before you rebuild it.
- Experimental
- Custom kernel required
Sources
- github.com/tonyd2wild/DeepSeek-V4.1-Flash-TP2-and-TP4-DGX-Sparks/blob/main/results/exl3tp4b-rep1/bench-exl3tp4b-rep1.md
- github.com/tonyd2wild/DeepSeek-V4.1-Flash-TP2-and-TP4-DGX-Sparks/blob/main/results/exl3tp4b/bench-exl3tp4b.md
- github.com/tonyd2wild/DeepSeek-V4.1-Flash-TP2-and-TP4-DGX-Sparks#four-sparks-on-exl3-the-context-lane
- github.com/tonyd2wild/DeepSeek-V4.1-Flash-TP2-and-TP4-DGX-Sparks/blob/main/docs/EXL3-TP3.md#tp4-lane-four-sparks-on-exl3-the-context-lane
Related recipes
8× DGX Spark
74tok/sEveryday
1 request, generation at 32K context (speed table, ring of eight), thinking on, high, with DSpark (per recipe); experimental profile Source
Intelligence40with thinkingno thinking 25
- Custom kernel required
- Experimental
4× RTX PRO 6000
73.1tok/sEveryday
1 request, prose prompt, context 32788, with DSPARK Source
Intelligence40with thinkingno thinking 25
4× DGX Spark
71tok/sEveryday
1 request, decode at 32K context (speed table, ring of four), thinking on, high Source
Intelligence40with thinkingno thinking 25
- Custom kernel required
4× DGX Spark
59.1tok/sEveryday
1 request, prose prompt, with DSpark (per recipe) Source
Intelligence40with thinkingno thinking 25
3× DGX Spark
51tok/sEveryday
1 request, prose prompt, with MTP/EAGLE (spekulativ, ~59 ms/step) (per recipe) Source
Intelligence40with thinkingno thinking 25
- Custom kernel required
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.