FR
Réserver un appel

Local AI · Recipe · 2× DGX Spark

DeepSeek-V4.1-Flash EXL3 2.0 bpw with TensorFold on 2× DGX Spark

DeepSeek-V4.1-Flash EXL3 2.0 bpw with TensorFold on 2× DGX Spark: 41.2 tok/s according to github.com (dataset as of Sep 29, 2026).

by sfxnz

41.2tok/sEveryday

1 request, prose prompt (512 tokens), with DSpark (per recipe)

Source: github.com
Intelligence (original model) · Artificial Analysis40with thinkingno thinking 25Original model's value. Heavily compressed (~2 bit), intelligence is likely much lower.

≤ 3 bit: quantization may cost quality

Instructions in the README Repo

Engine

Engine
TensorFold Fork sfxnz/TensorFold dsv41-recipe-engine4 @b86514a (v0.6.4 + deepseek_v41-Familie), TP=2, --parallel 4
Quantization
EXL3 2.0 bpw MCG (routed experts, Branch 2.0bpw-mcg-viterbi-lmhead-mxfp8 @982b704), Rest FP8 (32x32-Blockscales), lm_head MXFP8, DSpark-Draft MXFP4
Model family
DeepSeek-V4.1-Flash
Context
1,048,576
Parameters
Basis deepseek-ai/DeepSeek-V4.1-Flash: 763,205,315,794 Parameter (HF-API safetensors total)
Creator
sfxnz
GitHub stars
14
Repo updated
Oct 8, 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

  • Everyday (decode)41.2 tok/scontext 512-token prose · L.A.I.L-Zelle (measure_lail_prose.py: 512 Tokens, t=0.2, 10 Runs), s5 @903a1e8; c=1 bei b86514a innerhalb 0,5 %. bench_decode prose c=1 aktuell 56.3 (ignore_eos, 63 % nach EOS) Source 

Decode · parallel requests

Sums over all streams are marked; per request it is correspondingly less.

  • 4 parallel requests201 tok/sPeak (decode) · Structured, Aggregat c=4 (shipped --parallel 4), Median zweier Boots; c=1 structured 97.7 Source 

Prefill by context

  • 64K1,790 tok/sPrefill_cold.py, 65.536 Prompt-Tokens, Median 3 Prompts, s5 (Boots E/F 1800.2/1779.9) Source 

What you need

Hardware
2 × NVIDIA DGX Spark (GB10)
Engine
TensorFold Fork sfxnz/TensorFold dsv41-recipe-engine4 @b86514a (v0.6.4 + deepseek_v41-Familie), TP=2, --parallel 4
Context
1,048,576 tokens

Notes

What matters before you rebuild it.

  • ≤ 3 bit: quantization may cost qualityAt 3 bit and below the model may answer noticeably worse than the original. The intelligence number refers to the original.
  • Fork
  • Custom kernel required

Sources

Related recipes

2× DGX Spark

DeepSeek-V4.1-Flash EXL3 2.0 bpw with vLLM on 2× DGX Spark

by sfxnz

42.6tok/sEveryday

1 request, prose prompt, with DSpark (per recipe) Source 

Intelligence40with thinkingno thinking 25Original model's value. Heavily compressed (~2 bit), intelligence is likely much lower.
  • ≤ 3 bit: quantization may cost quality
  • Experimental
  • Custom kernel required

EXL3 2.0 bpw MCG (routed experts, tail-biting Viterbi re-encoded, scale refit), lm_head MXFP8, KV fp8 (8-GiB-Pin), DSpark-3vLLMWeights Repo Updated Sep 27, 2026

Details

← Back to overview

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.

See the AI workshopBack to overview