Link:UkisAI
Swift-1.5-GLM-5.3-Flash-NVFP4: Token-Efficient Reasoning on Blackwell GPUs
UkisAI has released an NVFP4 quantized version of GLM-5.3-Flash built with NVIDIA ModelOpt for Blackwell architectures. By penalising overthinking via post-training, the model cuts thinking tokens by 57.7% and speeds up inference with less than 1% accuracy loss. It caters to teams self-hosting agentic and coding workflows.
From X post: Post-Trained GLM-5.3-Flash Variants for Coding and Security