UnslothAI Releases Qwen3.6 MTP GGUF Models With Over 1.4x Faster Inference
UnslothAI发布Qwen3.6 MTP GGUF模型,实现推理速度大幅提升
Daniel Han released experimental Qwen3.6 MTP GGUF models, with the 27B model reaching 140 tokens/s on one GPU and the 35B-A3B version reaching 220 tokens/s, using two draft tokens for speculative decoding.
Why it matters: Unsloth founder Daniel Han released an experimental Qwen3.6 MTP GGUF that drafts two tokens per step, trading accuracy for over 1.4x faster inference.