📖 AI-Radar · 2026-W35

MTP lands in GLM-Air: 106B MoE speeds up on local GPUs

llama.cpp enables MTP for GLM-4.5-Air, a 106-billion-parameter MoE with only 12 billion active parameters. The change speeds up inference on memory-rich but compute-limited machines like Strix Halo, DGX Spark, and RTX 3090, and strengthens the fine-tuning ecosystem for creative writing.

24 Aug #Hardware #LLM On-Premise #Fine-Tuning