Local AI · Recipe · 1× DGX Spark

MiniCPM-V 4.7 35B-A3B EXL3 3.74 bpw with ExLlamaV3 on 1× DGX Spark

MiniCPM-V 4.7 35B-A3B EXL3 3.74 bpw with ExLlamaV3 on 1× DGX Spark: 60.3 tok/s according to github.com (dataset as of Nov 3, 2026).

by vcruz305

60.3tok/sEveryday

1 request, prose prompt (essay), context 2k, no drafter (per recipe)

Source: github.com

Intelligence (original model) · Artificial Analysisno independent value

Instructions in the README Repo

Engine: ExLlamaV3 Fork vcruz305 (master, gemessen dc778646 / 475461e9) + TabbyAPI Fork vcruz305
Quantization: EXL3 SAGE 3.74bpw (Kopf/Vision 8-bit), FP16-KV (8K) bzw. Q4-KV (262K)
Model family: MiniCPM-V-4.7-35B-A3B
Context: 262,144
Parameters: 35B MoE / 3B aktiv (lt. Modellname)
Creator: vcruz305
GitHub stars: 4
Repo updated: Oct 14, 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)60.3 tok/scontext 2K · Essay-Antwort (256 Tokens) auf synthetisches Dokument, 2.040 Prompt-Tokens, Q4-KV, max. 131K Kontext, target-only; bei vollem 262K-Kontext 45,03 Source
Peak (decode)69.5 tok/scontext 1K · Bestwert der Quelle ohne Drafter: 1.024-Token-Prompts (Inhalt nicht angegeben), 256 Tokens, greedy, p50 aus 7 Läufen, FP16-KV, 8K (Profil single); optional dSparK Breite 7: 94,70 Source

Prefill by context

262K1,056 tok/sKalt, 261.886-Token-Prompt, Q4-KV, TTFT 248,8 s Source

Time to first token

Shorter is better.

Prompt not stated948 mscontext 1K · 1.024-Token-Prompt, FP16-KV, p50, target-only Source

What you need

Hardware: 1 × NVIDIA DGX Spark (GB10)
Weights: Main weightsvcruz305/MiniCPM-V-4.7-35B-A3B-EXL3-SAGE-3.74bpw Apache-2.0
Baseopenbmb/MiniCPM-V-4.7-35B-A3B Apache-2.0
Engine: ExLlamaV3 Fork vcruz305 (master, gemessen dc778646 / 475461e9) + TabbyAPI Fork vcruz305
Context: 262,144 tokens

Notes

What matters before you rebuild it.

Custom kernel required
vision

Sources

Related recipes

1× DGX Spark

MiniCPM-V-4.7-35B-A3B EXL3 6.08 bpw with ExLlamaV3 on 1× DGX Spark (text, 262k context)

by r0b0tlab

59.4tok/sEveryday

1 request, chat prompt (20 prompts, thinking off), no drafting (default profile) Source

Intelligenceno independent value

Custom kernel required

EXL3 6.08 bpw (28,3 GB), FP16 KV; nur Textpfad (Vision-Tower wird von ExLlamaV3 nicht ausgeführt)ExLlamaV3 1.5.0 (Fork r0b0tlab/exllamav3-r0b0t, Branch minicpm-v4.7-dspark @ 5ddae7d12448), torch 2.14.0+cu130Weights Repo Updated Oct 14, 2026

Details

1× DGX Spark

Nemotron-3.5-Bolt 74B-A9B 24GB pack with TensorFold on 1× DGX Spark

by MiaAI-Lab

71.8tok/sEveryday

1 request, prose prompt (sampled, thinking off), with LiLiCorr (per recipe) Source

Intelligenceno independent value

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

24GB pack: IQ2_XXS experts (2.0625 bits), NVFP4 Mamba projections, Q8_0 attention/latent projections/lm_head; bf16 KVTensorFold tensorfold-native v1.1.6 (eea9dc9c) + recipe patches/ (Bolt lane backend, libbolt.so CUDA forward pass), LiLiCorr draftsWeights Repo Updated Nov 1, 2026

Details

1× DGX Spark

Qwen3.8 27B MLX 4-bit with TensorFold on 1× DGX Spark

by MiaAI-Lab

58.3tok/sEveryday

1 request, prose prompt, with DFlash2 (per recipe) Source

Intelligence34with thinking no thinking 20

Custom kernel required

MLX affine 4-bit (g64) + FP8 KV, FP8-PrefillTensorFold v0.6.0 (c464617) + 5 Patches, Image ghcr.io/miaai-lab/qwen3.8-27b-dgx-spark-tensorfold:v0.6.0, DFlash2-DraftsWeights Repo Updated Oct 1, 2026

Details

1× DGX Spark

Qwen3.8-Flash-Next NVFP4 (NVIDIA QAD, FP4-Local) with TensorFold on 1× DGX Spark

by MiaAI-Lab

55.6tok/sEveryday

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

Intelligence40with thinking

Custom kernel required
License unclear

NVIDIA FP4-QAD-Export (FP4-Local @b9ce18a6): Routed Experts NVFP4 (Gruppe 16), Attention/Shared Expert MXFP8, alle 512 Experten mit 10 pro Token; n-gram-Tabelle per pread von der SSD; FP8 KVTensorFold tensorfold-native (Zig) v1.1.4 (eab85828) + Patches 0001–0017, Image tensorfold-qwen38:v1.1.4-eab8582 (GHCR v1.1.4-eab8582-4baf16137ecd) auf nvcr.io/nvidia/pytorch:26.07-py3, MTP-DraftsWeights Repo Updated Oct 30, 2026

Details

1× DGX Spark

Qwen3.8-Flash-Next EXL3 3.05 bpw with ExLlamaV3 on 1× DGX Spark

by vcruz305

53.3tok/sEveryday

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

Intelligence40with thinking

Custom kernel required
License unclear

EXL3 3.05bpw (turboderp-Pack, h5, ngram5), 8-bit KV (K8/V8), n-gram im RAMExLlamaV3 Fork 24f0dece3 (≥1.6.0.post1, upstream 1.6.0) + TabbyAPI Fork 5a4f3efWeights Repo Updated Oct 14, 2026

Details

1× DGX Spark

Ornith-1.5-122B EXL3 4 bpw with TensorFold on 1× DGX Spark

by MiaAI-Lab

52.6tok/sEveryday

1 request, prose prompt, with DFlash2 (per recipe) Source

Intelligenceno independent value

Custom kernel required
Experimental

EXL3 EAQ v1 4 bpw (4.049 bpw trellis average per quant card), FP8 KVTensorFold tensorfold-native (Zig 0.17.0, pinned d1018b31 + recipe patches/ for EXL3 in the Qwen3.5-MoE engine), image tensorfold-ornith:zig-d1018b3, DFlash2 draftsWeights Repo Updated Oct 21, 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

MiniCPM-V-4.7-35B-A3B EXL3 6.08 bpw with ExLlamaV3 on 1× DGX Spark (text, 262k context)

Nemotron-3.5-Bolt 74B-A9B 24GB pack with TensorFold on 1× DGX Spark

Qwen3.8 27B MLX 4-bit with TensorFold on 1× DGX Spark

Qwen3.8-Flash-Next NVFP4 (NVIDIA QAD, FP4-Local) with TensorFold on 1× DGX Spark

Qwen3.8-Flash-Next EXL3 3.05 bpw with ExLlamaV3 on 1× DGX Spark

Ornith-1.5-122B EXL3 4 bpw with TensorFold on 1× DGX Spark

Local AI in your company?