Local AI · Recipe · 2× DGX Spark

GLM-5.3-Flash 320B-A18B EXL3 4 bpw with ExLlamaV3 on 2× DGX Spark

GLM-5.3-Flash 320B-A18B EXL3 4 bpw with ExLlamaV3 on 2× DGX Spark: 36.1 tok/s according to github.com (dataset as of Nov 3, 2026).

by MiaAI-Lab

36.1tok/sEveryday

1 request, prose prompt, with DFlash2, TR3 weights (per recipe)

Source: github.com

Intelligence (original model) · Artificial Analysis42with thinking

≤ 3 bit: quantization may cost quality

Instructions in the README Repo

Engine: TensorFold-basiertes Kit (ghcr.io/miaai-lab/glm-5.3-flash-2x-dgx-sparks:exl3-instanttensor), DFlash2 7 draft tokens; opt-in Speed-Boost-Profil (einmaliger ~25-min-Build auf dem Head)
Quantization: EXL3 4bpw (TR3) + FP8 KV (fp8_ds_mla)
Model family: GLM-5.3-Flash
Context: 850,000
Parameters: 320B MoE / 18B active (lt. tonyd2wild-README; Mia: 320B MoE)
Creator: MiaAI-Lab
GitHub stars: 750
Repo updated: Oct 18, 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)36.1 tok/sProsa, TP2-Default bis 04.10. (TR3-4bpw-Pack, byte-identisch mit brandonmusic Snapshot 5ab363a8), gemessen 2026-09-28, Mittelwert frischer Boots; opt-in Speed-Boost: 37,5 Source
Peak (decode)92.3 tok/sStructured output mit Speed-Boost-Profil (opt-in); Default: 83,0. Code 54,6/55,5 — gemessen 2026-09-28 Source

Prefill by context

32K1,511 tok/sCold prefill bei 32k (Default): 1.411 @8k / 1.511 @32k / 1.498 @96k — gemessen 2026-09-28 Source

What you need

Hardware: 2 × NVIDIA DGX Spark (GB10)
Weights: Main weightsbrandonmusic/GLM-5.3-Flash-tr3-4bpw other
Basezai-org/GLM-5.3-Flash MIT
Draft modelincoai/GLM-5.3-Flash-DFlash2 Non-commercial · CC-BY-NC-ND-4.0
otherbrandonmusic/GLM-5.3-Flash-tr3-4bpw other
Engine: TensorFold-basiertes Kit (ghcr.io/miaai-lab/glm-5.3-flash-2x-dgx-sparks:exl3-instanttensor), DFlash2 7 draft tokens; opt-in Speed-Boost-Profil (einmaliger ~25-min-Build auf dem Head)
Context: 850,000 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
Non-commercial
License unclear

Sources

Related recipes

4× DGX Spark

GLM-5.3-Flash NVFP4/MXFP8 CSF with SparkRing on 4× DGX Spark

by fujitsupolycom

68tok/sEveryday

1 request, generation at 32K context (speed table; measured only as a path of four Sparks inside a ring of eight, not as a ring of four), thinking always on (measured before the profile default high, i.e. template default max), with MTP3 (per recipe) Source

Intelligence42with 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

4× DGX Spark

GLM-5.3-Flash EXL3 4bpw with TensorFold on 4× DGX Spark

by tonyd2wild

65.4tok/sEveryday

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

Intelligence42with thinking

Custom kernel required
Non-commercial
License unclear

EXL3 4bpw routed experts (mcg) + q4 dense (at load) + FP8 KVTensorFold v0.6.0 + MiaAI-Lab GLM-5.3-Flash patch series (engine-multi, + tonyd2wild changes)Weights Repo Updated Oct 12, 2026

Details

2× DGX Spark

GLM-5.3-Flash EXL3 with TensorFold on 2× DGX Spark (1M context, 4 streams)

by MiaAI-Lab

60.4tok/sEveryday

Prose decode, 1 request; quality (FP8 KV, 4-bit dense): GSM8K 98.8% / HumanEval 95.7% / HumanEval+ & MBPP+ 86.5% Source

Intelligence42with thinking

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

EXL3/TR3 4bpw routed experts + 4-bit dense, BF16 elsewhere, FP8 KVTensorFold v0.6.0 + Recipe v1.10 (README: 96 Patches; gepinntes Image v0.6.0-e66c657a433d mit 107 Patches: DFlash2+Copy-Drafts, 4-bit dense, FP8-KV, RoCE-One-Shot-All-Gather, Shared KV-Pool, optionaler Spill-Tier, Vision, Tool-Calling, Anthropic /v1/messages; PARALLEL=8 opt-in)Weights Repo Updated Oct 25, 2026

Details

2× DGX Spark

GLM-5.3-Flash EXL3 4bpw with TensorFold on 2× DGX Spark

by tonyd2wild

44tok/sEveryday

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

Intelligence42with thinking

Custom kernel required
Non-commercial
License unclear

EXL3 4bpw routed experts (mcg) + q4 dense (at load) + FP8 KVTensorFold v0.6.0 + MiaAI-Lab GLM-5.3-Flash patch series (engine-multi, + tonyd2wild changes)Weights Repo Updated Oct 12, 2026

Details

4× DGX Spark

GLM-5.3-Flash 320B-A18B NVFP4 with vLLM on 4× DGX Spark

by tonyd2wild

40.8tok/sEveryday

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

Intelligence42with thinking

Custom kernel required
Non-commercial

NVFP4 weights + FP8 KV (Lane B: NVFP4 KV)vLLM 0.28.0Weights Repo Updated Oct 10, 2026

Details

3× DGX Spark

GLM-5.3-Flash 320B EXL3 4 bpw with vLLM on 3× DGX Spark

by MiaAI-Lab

40.1tok/sEveryday

1 request, prose prompt, with DFlash2, TR3 weights (per recipe) Source

Intelligence42with thinking

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

EXL3/TR3 4 bpw (routed experts), FP8 KV (fp8_ds_mla)vLLMWeights Repo Updated Oct 18, 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

What you need

Notes

Sources

Related recipes

GLM-5.3-Flash NVFP4/MXFP8 CSF with SparkRing on 4× DGX Spark

GLM-5.3-Flash EXL3 4bpw with TensorFold on 4× DGX Spark

GLM-5.3-Flash EXL3 with TensorFold on 2× DGX Spark (1M context, 4 streams)

GLM-5.3-Flash EXL3 4bpw with TensorFold on 2× DGX Spark

GLM-5.3-Flash 320B-A18B NVFP4 with vLLM on 4× DGX Spark

GLM-5.3-Flash 320B EXL3 4 bpw with vLLM on 3× DGX Spark

Local AI in your company?