GLM-5.3-Flash

Overview

The first natively multimodal GLM-5 - 320B total with only 18B active, hybrid linear + sparse attention for cheap long-context serving, and a 1M context window. Approaches Opus 4.8 on coding and agentic benchmarks at a tenth of GLM-5.3's serving cost. Official FP8 build for vLLM and SGLang.

Good for

• agentic
• chat
• coding

Recommended hardware

≈ 384 GB VRAM to run this 321B model at the FP8 quantization.

Needs a multi-GPU server node (e.g. 8x 80 GB H100, or 4x 192 GB)

Links

• View on HuggingFace
• Run on Infersec