DSpark baseline for Qwen3-8B

This repository holds the DSpark draft model that the JetSpec++ paper compares against, for the target Qwen/Qwen3-8B. The JetSpec++ authors trained it with the DSpark method.

Contents

  • config.json, dspark.py, dspark_heads.py, model.safetensors: the draft model in bf16.
  • manifest.json: the sha256 and size of every file.

The JetSpec++ serving engine registers these files as blockhead800k_checkpoint and checks the sha256 of each file. bench/jetspec_plus_plus/table3.py runs this model as the DSpark arm of Table 3.

Results

The 8B DSpark numbers in Table 3 of the paper were measured with these files.

Downloads last month
51
Safetensors
Model size
2B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for JetSpec/dspark-qwen3-8b-800k

Finetuned
Qwen/Qwen3-8B
Finetuned
(2,303)
this model

Papers for JetSpec/dspark-qwen3-8b-800k