Quantized GLM-5.3-Flash lands on Hugging Face for DGX Spark
A community NVFP4 quantization of Zai's GLM-5.3-Flash is tuned to run on Nvidia's GB10-based DGX Spark desktop box.
The facts
- Uploader autotrust published GLM5.3-Flash-E224-DGX-Spark, a quantized derivative of zai-org/GLM-5.3-Flash
- The model uses NVFP4 quantization via Nvidia's ModelOpt, targets the GB10 (Blackwell) chip in Nvidia's DGX Spark, and is described as expert-pruned
- Total parameter count is listed at roughly 127.6 billion, with most weights (118.6B) stored in 8-bit precision and a smaller BF16 portion (9.0B)
- Repo files total about 151.5GB sharded across safetensors, licensed MIT, tagged for vllm serving, and the model card lists 5,422 downloads and 521 likes as of its listing
A Hugging Face user going by autotrust has published GLM5.3-Flash-E224-DGX-Spark. It is a quantized build of Zai's GLM-5.3-Flash model. The target is Nvidia's DGX Spark desktop AI box.
The repo carries four tags: nvfp4, modelopt, dgx-spark, gb10 and blackwell. These point to Nvidia's four-bit floating point format, its ModelOpt quantization toolkit, and the GB10 Blackwell chip that powers DGX Spark.
The model card lists it as expert-pruned. That is a mixture-of-experts technique that trims the number of experts to cut memory and compute needs. It is marked image-text-to-text and conversational, and served through the vllm inference engine.
The safetensors metadata shows roughly 127.6 billion total parameters. Of those, 118.6 billion are stored in 8-bit precision and 9.0 billion are left in BF16. The full sharded weight set runs to about 151.5GB on disk.
The repo is licensed MIT and built from the base model zai-org/GLM-5.3-Flash. As of the listing it shows 5,422 downloads and 521 likes.
The source text gives no benchmark numbers and no stated context window. Nor does it say how the pruning or quantization affects output quality against the unquantized GLM-5.3-Flash base.
Why it matters
A 150GB quantized MoE model built for one desktop accelerator shows the community is now packaging large open models specifically for DGX Spark, not treating it as an afterthought. Anyone evaluating on-prem inference hardware should watch whether vendor-specific quant builds like this become standard alongside full-precision and generic-quantized releases.
Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.