ModelQwen3-4B-Inst-2507-MTP

Qwen3-4B-Inst-2507-MTP

A 4-billion-parameter instruction-tuned model version in Alibaba's Qwen series that supports multi-token prediction.

Timeline (last 90 days)

Oct 21

Across five draft-target pairs built from Llama-3.1-8B-Instruct with Medusa heads, Meta's MTP models, and Qwen3-4B-Inst-2507-MTP, AMTP achieves up to 4.1x speedup over standard decoding.

Unverified50%

All Facts (1)

Unverified

Across five draft-target pairs built from Llama-3.1-8B-Instruct with Medusa heads, Meta's MTP models, and Qwen3-4B-Inst-2507-MTP, AMTP achieves up to 4.1x speedup over standard decoding.

50%

Source Articles

Adaptive Multi-Token Prediction: How Does the Speculative Horizon Dynamically Adjust for Inference Acceleration?Oct 21