JetSpec/jetspec-plus-plus-qwen3-8b

JetSpec++ for Qwen3-8B

This repository holds the JetSpec++ draft model, adapter and calibration for the target Qwen/Qwen3-8B (revision b968826d).

• Paper: JetSpec++
• Code: hao-ai-lab/JetSpec

Contents

• config.json, dflash.py, model.safetensors: the drafter, 1.05B parameters in bf16. It uses the same format as JetSpec/jetspec-qwen3-8b and loads with AutoModel.from_pretrained(..., trust_remote_code=True).
• adapter/: the adapter (head.safetensors, converter.safetensors, config.json).
• calibration/thresholds.json: the serving thresholds for block size 16.
• calibration/decoder/manifest.json: the decoder is a slice of the target's last layers. This file pins the target files that the engine cuts it from.
• manifest.json: the sha256 and size of every file.

The engine takes the drafter's embedding and LM head from the target, so this repository does not include them.

Results

The 8B JetSpec++ numbers in the paper (Tables 1, 3, 5 and 16) were measured with these files.