Reddit
llama.cpp Merged Multi-Token Prediction for GLM-4.5-Air, Giving a Measured 1.17x to 1.20x Speedup on a 106B MoE
PR 26534 landed in ggml-org/llama.cpp on 2026-08-23 after three weeks in review, implementing graph_mtp for the glm4moe architecture. Testing across multiple Hugging Face GGUFs, including community fine-tunes, showed a consistent 1.17x to 1.20x speedup with --spec-draft-n-max 1, and full GLM-4.5 gained about 1.07x with --n-cpu-moe 30. The converter now supports --no-mtp, --mtp and combined GGUFs, and for weights without a NextN block there is a separate 'GLM-4.5-Air-MTP-GGUF' file you can drop in as an external draft. GLM-4.5-Air is 106B with 12B active, which makes it a fit for large-memory low-compute boxes like Strix Halo and DGX Spark.
Source
↳ Follow the thread