Local AI · Recipe · 1× DGX Spark

GLM-5.3-Flash 320B UD-IQ3_XXS with llama.cpp on 1× DGX Spark

GLM-5.3-Flash 320B UD-IQ3_XXS with llama.cpp on 1× DGX Spark: 20.8 tok/s according to github.com (dataset as of Nov 3, 2026).

by Weschera

20.8tok/sMixed

1 request, mixed prompt set, with MTP

Source: github.com
Intelligence (original model) · Artificial Analysisno independent value

≤ 3 bit: quantization may cost quality

Engine

Engine
llama.cpp
Quantization
Unsloth Dynamic UD-IQ3_XXS GGUF (120,4 GB, 4 Shards); KV q8_0
Model family
GLM-5.3-Flash
Context
65,536
Parameters
320B / A18B MoE (README)
Creator
Weschera
GitHub stars
5
Repo updated
Sep 4, 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

  • Mixed (decode)20.8 tok/sMedian p10–p90 11.6–25.1, 50 Requests / 26K Tokens, thinking OFF, aus der Benchmark-Run; MTP n=2 Source 
  • Peak (decode)20.8 tok/sMedian p10–p90 11.6–25.1, 50 Requests / 26K Tokens, thinking OFF, aus der Benchmark-Run; MTP n=2 Source 

Prefill by context

  • 7K270 tok/s6.981-token-Prompt in 26 s, README '~270 tok/s' → approx Source 

What you need

Hardware
1 × NVIDIA DGX Spark (GB10)
Engine
llama.cpp
Context
65,536 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.

Sources

Related recipes

4× DGX Spark
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
2× DGX Spark
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

← 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.