Open-weight catalog
Models, without a fake universal leaderboard.
Quality bands appear only when comparable evidence supports them. Provider claims, repository-agent results, and context evaluations remain separate records.
Qwen3-Coder 30B-A3B Instruct
Strong coding-oriented choice based on model-specific evidence; benchmark scores are shown individually rather than merged.
Qwen2.5-Coder 32B Instruct
Mature coding model with disclosed architecture; individual benchmark evidence remains context-specific.
Devstral Small 2505
Provider reports 46.8% SWE-bench Verified with OpenHands; retained as a provider claim, not a universal quality score.
DeepSeek Coder V2 Lite Instruct
Useful lightweight coding MoE; shown without a composite quality rank.
Qwen2.5-Coder 7B Instruct
A responsive entry-level coding model; capability is not directly comparable with larger agentic models.
Llama 3.3 70B Instruct
Included as a general high-capability comparison, not claimed to be a coding-specialist winner.