Magicoder 6.7B
The synthetic-data pioneer — OSS-Instruct showed LLM-generated code data could rival the giants.
About
Magicoder-S-DS-6.7B (early 2024, UIUC) is academically famous: the OSS-Instruct paper showed that a 6.7B model trained on LLM-synthesized code instruction data could match much larger models — pioneering the synthetic-data recipe that the entire field then adopted.
Licensing: the Hugging Face checkpoint carries a custom "other" tag — verify terms. It's a DeepSeek-Coder-6.7B fine-tune, so the base license lineage matters too. 4.5k Hugging Face downloads. A 2024 6.7B model superseded by Qwen2.5-Coder-7B, but the OSS-Instruct methodology influence is enormous — this is the paper that changed how code models get their data. For research on synthetic data; for deployment, use Qwen2.5-Coder.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-claimed (paper): matched larger models on HumanEval — treat as provider-claimed; paper-reported on standard benchmarks (HumanEval). Enormous methodological influence.
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 (metrics tier). The OSS-Instruct paper changed the field's data recipe — enormous methodological influence. Docked five points: (1) custom checkpoint license, verify terms; (2) 2024 6.7B, superseded by Qwen2.5-Coder-7B; (3) modest adoption (4.5k HF downloads); (4) 16K context; (5) academic artifact, no product support. Read the paper; deploy Qwen.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.7/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 4.6/10
- Manual review · sanity read + red-flag check
- 5.0/10
- · HF downloads 4,518 (peer-normalized)
- · released 1002 days ago
- · manual review: OSS-Instruct changed the field's data recipe; academic artifact; superseded by Qwen2.5-Coder-7B; license: custom checkpoint license (other tag) — verify; DeepSeek-Coder base lineage