JetSpec/jetspec-plus-plus-qwen3p8-27b

JetSpec++ for Qwen3.8-27B

This repository holds the JetSpec++ draft model, adapter and calibration for the target Qwen/Qwen3.8-27B (revision 1d4bf0f2).

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

Contents

• config.json, dflash.py, model.safetensors: the drafter in bf16, in the same format as the other JetSpec drafters. It loads with AutoModel.from_pretrained(..., trust_remote_code=True).
• adapter/: the adapter (head.safetensors, converter.safetensors, config.json).
• calibration/thresholds.json: the serving thresholds.
• 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 paper's 27B numbers (Tables 2 and 3) were measured with an earlier adapter, and that adapter needed a system message in the prompt. This adapter is retrained so that it does not need one. With a system message, the two adapters give the same acceptance length within run-to-run noise. Without a system message, this adapter accepts more tokens per round.