Community · Nov 4, 2026, 21:47 UTC

Inference engine serves GLM-5.3-Flash on AWS Trainium2 using NKI megakernels

Open-source engine runs the 320B MoE GLM-5.3-Flash with 1M-token context on one trn2.48xlarge. Batch-8 decode step fell from 32.2 ms to 24.2 ms, and up to 28.9 agent requests a minute were measured on AgentX, self-run and not official.

• LabAmazon
• LabZhipu
• ModelGLM-5.3-Flash
• Topicchips
• Topiccoding
• Topicopen-weights
• Topicresearch