DiffuCoder
Apple and HKU's 7B masked-diffusion code model, adapted from Qwen2.5-Coder and tuned with coupled-GRPO; research license.
About
DiffuCoder-7B-Base is Apple's entry into diffusion language models, built with the University of Hong Kong and released in July 2025. It starts from Qwen2.5-Coder-7B and converts it to a masked-diffusion model through a four-stage pipeline: adaptation pretraining on 130B code tokens, mid-training, instruction tuning on 436K samples, and a final reinforcement-learning stage using coupled-GRPO on hard coding problems. The modeling code reuses the Dream architecture, and Apple published the paper (arXiv:2506.20639) with full training code on GitHub.
The interesting claim is about decoding: unlike most open diffusion LMs, which fall back to semi-autoregressive decoding that throws away the parallelism advantage, DiffuCoder is trained to refine code in parallel the way a programmer revises a file. Apple positions it as a strong baseline on HumanEval, MBPP, and BigCodeBench, though the model card states this qualitatively without a numeric table. The catch is the license: Apple Machine Learning Research terms, which are research-oriented rather than a permissive open license, so check the terms before shipping anything commercial on top of it.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Apple reports strong baseline performance on HumanEval, MBPP and BigCodeBench for DiffuCoder-7B-Base (model-card statement, no numeric table published)
- Paper (arXiv:2506.20639) includes HumanEval/MBPP ablations; SOTA-level numeric claims not reproduced here
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 5/10 (verified tier, metrics-v1). Evidence: HF repo apple/DiffuCoder-7B-Base created Jul 1, 2025 (HF API); paper arXiv:2506.20639 with Apple + HKU authors; training code at github.com/apple-aiml-research/ml-diffucoder. Weights downloadable. Strengths: serious engineering (four-stage pipeline, coupled-GRPO), full code and paper published, Apple research backing. Docked for: (1) Apple ML Research license is research-oriented, not a permissive open license; (2) benchmark claims are qualitative on the model card with no numeric table; (3) tiny adoption (303 downloads) and no sign of continued maintenance. License: apple-amlr (Apple Machine Learning Research, research terms).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 4.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 5.0/10
- · HF repo apple/DiffuCoder-7B-Base created 2025-07-01; 303 downloads, 30 likes (HF API Oct 2026)
- · Paper arXiv:2506.20639 (Apple + HKU); training code at github.com/apple-aiml-research/ml-diffucoder
- · Adapted from Qwen2.5-Coder-7B; four-stage pipeline incl. coupled-GRPO (HF model card)
- · Apple ML Research license is research-oriented, not permissive open
- · Benchmark claims qualitative on model card; no numeric table published