distill-WeDLM-KL is a 0.6B diffusion language model distilled from WeDLM-8B-Instruct (8B dense) into the Qwen3-0.6B-diffusion-bd3lm-v0.1 student in the Shared-Tokenizer (Pipeline B) of the TIDE framework. Forward token-level KL distillation baseline.

Model Overview

Method: TIDE — Reverse CALM / TIDAL / CompDemo (cross-architecture distillation for diffusion LMs)
Framework: TIDE / dLLM
Student (initialization): Qwen3-0.6B-diffusion-bd3lm-v0.1 (BD3LM, block_size=32)
Teacher: tencent/WeDLM-8B-Instruct
Distillation mode: --distill_mode kl_aligned
Datasets: tulu-3-sft-mixture, smoltalk, opc-sft-stage1 and opc-sft-stage2 — same composition as the Qwen3-0.6B-diffusion-bd3lm-v0.1 base. Pre-tokenized for this teacher in TIDE-dllm/distill_wedlm_sft.

Installation

pip install torch transformers accelerate

Quick Start

[!NOTE] This checkpoint is fully compatible with the BD3LM generate(...) routine published with dllm-hub/Qwen3-0.6B-diffusion-bd3lm-v0.1 — only the model name changes.

python

import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer

repo = "TIDE-dllm/distill-WeDLM-KL"
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForMaskedLM.from_pretrained(
    repo, dtype=torch.bfloat16, trust_remote_code=True,
).to(device).eval()
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)

prompts = [
    [
        {"role": "system", "content": "You are a helpful AI assistant."},
        {"role": "user", "content": "Implement a DFS traversal in Python with clear inline comments."},
    ],
]
encoded = [tokenizer.apply_chat_template(m, add_generation_prompt=True, tokenize=True, enable_thinking=False) for m in prompts]
# ... use the same `generate()` function as in dllm-hub/Qwen3-0.6B-diffusion-bd3lm-v0.1.

Command-Line Interface

For an interactive demo (visualised iterative denoising), use the script in the TIDE / dLLM repo:

shell

python -u examples/a2d/bd3lm/chat.py \
    --model_name_or_path TIDE-dllm/distill-WeDLM-KL \
    --chat_template True --block_size 32 --remasking low_confidence \
    --steps 256 --max_new_tokens 256

Reproducing this checkpoint

shell

git clone https://github.com/PKU-YuanGroup/TIDE && cd TIDE
pip install -e . && git submodule update --init --recursive
pip install -e "lm-evaluation-harness[ifeval,math]" && pip install -e "tokenkit[full]"

# Download the pre-tokenized SFT mixture for this teacher
huggingface-cli download TIDE-dllm/distill_wedlm_sft --repo-type dataset \
    --local-dir data/distill_wedlm_sft

bash scripts/distill_wedlm.sh \
    --data_path data/distill_wedlm_sft \
    --distill_mode kl_aligned \
    --num_gpus 8

Citation

bibtex

@misc{zhang2026turningtidecrossarchitecturedistillation,
      title={Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models},
      author={Gongbo Zhang and Wen Wang and Ye Tian and Li Yuan},
      year={2026},
      eprint={2604.26951},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.26951},
}