Unsloth Ships Multi-Token Prediction GGUFs for Qwen3.6 — 1.5-2x Faster Local Inference with 75% Acceptance Rate

Hugging Face / Unsloth·medium signal

Unsloth released GGUF quantizations of Qwen3.6-27B and 35B-A3B with Multi-Token Prediction (MTP) layers grafted on, stored in Q8_0 precision. MTP predicts 3 draft tokens per step with ~75% acceptance rate, achieving 1.5-2x faster generation over baseline autoregressive decoding. Requires a custom llama.cpp build from PR #22673. Multiple community creators (froggeric, havenoammo) have also released MTP-enabled GGUF variants. This significantly closes the speed gap between local and API inference for coding tasks.