Local AI · Recipe · 2× DGX Spark

MiniMax-M3.1-Flash EXL3 2.15 bpw with vLLM (EXL3 plugin) on 2× DGX Spark

MiniMax-M3.1-Flash EXL3 2.15 bpw with vLLM (EXL3 plugin) on 2× DGX Spark: 19.8 tok/s according to github.com (dataset as of Nov 3, 2026).

by tonyd2wild

19.8tok/sEveryday

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

Source: github.com

Intelligence (original model) · Artificial Analysisno independent value

≤ 3 bit: quantization may cost quality

Instructions in the README Repo

Engine: vLLM + vllm-exl3 plugin (vcruz305 image/forks, pin 41fdb7d) + 2 TP2 patches
Quantization: EXL3 SAGE 2.15 bpw + q4 KV (fp8 indexer/draft KV)
Model family: MiniMax-M3.1-Flash
Context: 300,000
Parameters: 428B MoE / 23B aktiv
Creator: tonyd2wild
GitHub stars: 3
Repo updated: Oct 24, 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)19.8 tok/sProse, Single Stream, Median von 2, temp 0, DSpark (3 Draft-Tokens) Source
Mixed (decode)31.7 tok/sGemischte reale Prompts (8 Typen rotiert, ohne Zählen), C1, Decode pro Stream, Median von 2 Runden Source
Peak (decode)39.4 tok/sCount to 300 (repetitiv), Single Stream, Median von 2 Source

Prefill by context

97.51527 tok/sCold (salted prompt, 1 Output-Token), TTFT 185.0 s Source

Time to first token

Shorter is better.

Prompt not stated670 msProse, Single Stream Source

What you need

Hardware: 2 × NVIDIA DGX Spark (GB10)
Weights: Main weightsvcruz305/MiniMax-M3.1-Flash-EXL3-SAGE-2.15bpw other
BaseMiniMaxAI/MiniMax-M3.1-Flash other
Engine: vLLM + vllm-exl3 plugin (vcruz305 image/forks, pin 41fdb7d) + 2 TP2 patches
Context: 300,000 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.
Custom kernel required
License unclear

Sources

Related recipes

1× DGX Spark

MiniMax-M3.1-Flash EXL3 2.15 bpw with vLLM on 1× DGX Spark

by vcruz305

24.4tok/sEveryday

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

Intelligenceno independent value

≤ 3 bit: quantization may cost quality
Custom kernel required
License unclear

EXL3 SAGE 2.15bpw, q4-KV (nvfp4-Layout), fp8 Indexer-/Draft-KVvLLM (vllm/vllm-openai:minimax-m31-flash, Digest 86f582da) + vllm-exl3 Fork 41fdb7d + exllamav3 Fork fdd7736, DSpark k=4Weights Repo Updated Oct 24, 2026

Details

4× DGX Spark

MiniMax-M3.1-Flash NVFP4/MXFP8 with vLLM on 4× DGX Spark

by tonyd2wild

22.2tok/sEveryday

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

Intelligenceno independent value

License unclear

NVFP4 experts (Marlin W4A16 on GB10) + MXFP8 attention/dense, FP8 KV (official checkpoint)vLLM 0.1.dev23341+g8744dd2fb (vllm/vllm-openai:minimax-m31-flash-arm64-cu134)Weights Repo Updated Oct 23, 2026

Details

2× DGX Spark

Qwen3.8-Flash-Next 125B INT4-AutoRound with TensorFold on 2× DGX Spark (1M context)

by MiaAI-Lab

89.3tok/sEveryday

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

Intelligenceno independent value

Custom kernel required
License unclear

INT4-AutoRound (GPTQ int4, Top-5-Experten) + FP8 n-gram-Tabelle, FP8 KVTensorFold tensorfold-native (Zig, zig-flashnext db281878 + patches), Image ghcr.io/miaai-lab/qwen3.8-flash-dual-dgx-sparks-tensorfold:zig-db28187-f3e1300f4f62, TP=2 über CX7/RoCE, MTP + Copy-Drafts (QUANT=int4ar)Weights Repo Updated Oct 26, 2026

Details

2× DGX Spark

Qwen3.8-Flash-Next INT4-AutoRound (v5.2) with vLLM on 2× DGX Spark

by myllmbox

74tok/sEveryday

1 request, prose prompt (long story), thinking off, with MTP Source

Intelligence40with thinking

Custom kernel required
License unclear

INT4-AutoRound (azampatti): int4-Experten, fp8-Seitenschichten; n-gram-Tabelle beim Laden fp8 → NVFP4, halb pro Box auf der GPU; KV-Pin 29G/Box (1.829.182 Tokens)vLLM 0.30 (Image myllmbox/qwen38-flash-next-cluster-vllm:v5.2; RecoverSSM, dynamischer Draft-Depth 3–7 in Full-CUDA-Graphs, sampled Drafts + Block-Verifikation, One-Shot-RoCE-All-Reduce, Dual-Rail, TP=2, 64 Seats) — Kit v5.2Weights Repo Updated Oct 29, 2026

Details

2× DGX Spark

Qwen3.8-Flash-Next MLX 4-bit with TensorFold on 2× DGX Spark

by sfxnz

69.6tok/sEveryday

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

Intelligence40with thinking

License unclear

MLX affine 4-bit (group 32) auf allen Linears inkl. n-gram-Tabellen und MTP-Head, KV bf16TensorFold v0.6.2 (56e2e3e), TP=2, Image tf-qwen38-flashnext:0.6.2 (nvcr.io/nvidia/pytorch:26.07-py3)Weights Repo Updated Oct 2, 2026

Details

2× DGX Spark

Qwen3.8-Flash-Next NVFP4/MXFP8 CSF with SparkRing on 2× DGX Spark

by fujitsupolycom

68tok/sEveryday

1 request, generation at 32K context (speed table, pair), thinking on, xhigh, with MTP3 (per recipe) Source

Intelligence40with thinking

Custom kernel required
License unclear

NVFP4/MXFP8 CSF QAD (Local Inference Lab)vLLM (local-inference-lab Fork) via SparkRing (SIRCL-Collectives, switchless RDMA ueber ConnectX-7), Image Release 2026.10.2Weights Repo Updated Nov 2, 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 workshop Back to overview

Engine

Measurements

Decode · 1 request

Prefill by context

Time to first token

What you need

Notes

Sources

Related recipes

MiniMax-M3.1-Flash EXL3 2.15 bpw with vLLM on 1× DGX Spark

MiniMax-M3.1-Flash NVFP4/MXFP8 with vLLM on 4× DGX Spark

Qwen3.8-Flash-Next 125B INT4-AutoRound with TensorFold on 2× DGX Spark (1M context)

Qwen3.8-Flash-Next INT4-AutoRound (v5.2) with vLLM on 2× DGX Spark

Qwen3.8-Flash-Next MLX 4-bit with TensorFold on 2× DGX Spark

Qwen3.8-Flash-Next NVFP4/MXFP8 CSF with SparkRing on 2× DGX Spark

Local AI in your company?