---
title: "Red Hat Shrinks GLM-5.3-Flash to 4-Bit for NVIDIA Blackwell GPUs"
subtitle: "Red Hat AI shipped an NVFP4 quantization of Z.ai's 320B GLM-5.3-Flash, shrinking the model to run on vLLM with FP4 activations while holding reasoning benchmarks near the original."
url: "https://alphasignal.ai/news/red-hat-shrinks-glm-5-3-flash-to-4-bit-for-nvidia-blackwell-gpus"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-08-27T02:13:22.000Z"
updated: "2026-09-12T20:04:38.391Z"
type: "Model"
topics:
  - "Gpus"
  - "Infra"
subtopics:
  - "Quantization"
  - "Inference Optimization"
  - "Model Serving"
upvotes: 452
source_url: "https://huggingface.co/RedHatAI/GLM-5.3-Flash-NVFP4"
comments: 0
truncated: true
---

# Red Hat Shrinks GLM-5.3-Flash to 4-Bit for NVIDIA Blackwell GPUs

> Red Hat AI shipped an NVFP4 quantization of Z.ai's 320B GLM-5.3-Flash, shrinking the model to run on vLLM with FP4 activations while holding reasoning benchmarks near the original.

Red Hat AI has published an [NVFP4 checkpoint](https://huggingface.co/RedHatAI/GLM-5.3-Flash-NVFP4) of Z.ai’s MIT-licensed GLM-5.3-Flash. The release compresses the model’s transformer layers to a 4-bit floating-point format designed for NVIDIA Blackwell GPUs, with serving support through vLLM and SGLang.

GLM-5.3-Flash is a multimodal mixture-of-experts model with 320 billion total parameters and about 18 billion active for each token. Activating a smaller subset reduces per-token computation, while serving still requires access to the complete 320B-parameter checkpoint.

Z.ai reports that GLM-5.3-Flash outperforms GLM-5.2 across its benchmark suite at one-tenth the inference price and approaches Claude Opus 4.8 on coding and agentic evaluations. The architecture combines linear attention for local state tracking with indexed sparse attention for retrieving global context. Its advertised context window reaches one million tokens.

## Four bits, selectively applied

NVFP4 is NVIDIA’s 4-bit floating-point format for Blackwell Tensor Cores. Red Hat quantized weights and activations inside the transformer’s linear operators, using groups of 16 values with local scaling. The vision tower, token embeddings, and output head retain their original precision to limit quality loss in sensitive parts of the network.

Red Hat created the checkpoint with [LLM Compressor](https://github.com/vllm-project/llm-compressor), the vLLM project’s open-source quantization toolkit. The files use the `compressed-tensors` format and contain F32, BF16, U8, and F8\_E4M3 tensors. U8 tensors hold packed 4-bit values, FP8 tensors store scaling data, and higher-precision tensors cover the modules excluded from quantization.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [RedHatAI/GLM-5.3-Flash-NVFP4 on Hugging Face](https://huggingface.co/RedHatAI/GLM-5.3-Flash-NVFP4)
- [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash)
- [GLM-5.3-Flash: The Stealth Model That Became the Talk of the Timeline](https://www.gmicloud.ai/en/blog/glm-53-flash-the-stealth-model-that-became-the-talk-of-the-timeline)
- [GLM-5.3-Flash: Z.ai Reveals Ox Alpha](https://dev.to/jamilxt/glm-53-flash-zai-reveals-ox-alpha-was-its-open-multimodal-model-51b7)
- [Unsloth GLM-5.3-Flash documentation](https://unsloth.ai/docs/models/glm-5.3-flash)
