NVIDIA Releases GLM-5.3-Flash-FP4-Local Model
NVIDIA has launched the GLM-5.3-Flash-FP4-Local model, a quantized version of ZAI's GLM-5.3-Flash model, optimized for local AI hardware.
The NVIDIA GLM-5.3-Flash-FP4-Local model, a quantized version of ZAI’s GLM-5.3-Flash, is designed for reasoning, coding, and agentic tasks, utilizing a hybrid sparse and linear attention architecture to support a long context. This model is multimodal, handling text, image, and video inputs, and is optimized for local AI hardware, requiring approximately 166 GiB of vRAM. It is available for commercial or non-commercial use under the NVIDIA Open Model Agreement, with the GLM-5.3-Flash base model governed by the MIT License. You can run the model using vLLM on NVIDIA RTX PRO 6000 Blackwell or DGX Spark hardware, with specific containers provided for deployment. According to NVIDIA, the model achieves faster training and inference times on NVIDIA GPU-accelerated systems compared to CPU-only solutions. The model’s performance has been evaluated using GPQA Diamond and ToolEvalBench benchmarks, with NVIDIA claiming precision rates of 91.3% and 92.4% respectively for the FP4-Local configuration.