The best coding models, according to hackers
15 coding models recommended in Hacker News comments over the past year, ranked by how many separate people recommend each one.
- Most recommendedCursor Composer214 people recommend it
- Built to lastQwen3-Coderowners report a median of 0.17 years
- Least regretMeta Muse Glimmer 30B70 recommend, 18 criticize
All coding models
- Cursor ComposerCursor's in-house coding model is recommended for being very fast and cheap with solid results on planned or simple tasks, though critics say it trails frontier models on complex work.214recommend
- Qwen3-CoderA free open-weight coding model that runs locally at usable speed and handles routine coding and tool calling well, though it trails Claude on harder work and can loop.165recommend
- GPT-5.3-CodexA careful OpenAI coding model praised for long tasks, orchestration, verification and code review at good value; critics report silent routing to weaker models and robotic, hard-to-follow explanations.135recommend
- Meta Muse Glimmer 30BPraised as a terse, efficient local coding model that runs well on consumer hardware; critics say it trails Qwen 3.6 on benchmarks and distrust Meta.70recommend
- GPT-5.2-CodexPeople recommend it for strong coding, bug finding and code review. Common caveat: it is not clearly the best across the board and makes design errors.54recommend
- Laguna S 2.1Recommended as a fast, capable mixture-of-experts coding model that runs on home hardware; critics say it is weak outside coding and early releases had bugs.43recommend
- Devstral 2People recommend it as a strong open-weight coding model that runs locally on consumer hardware and keeps code off the cloud, though results vary by task.40recommend
- GPT-5.3-Codex-SparkRecommended for very fast coding-model output, around 1000 tokens per second, useful for quick edit and search loops. Caveats: small context window and weaker reasoning than regular Codex 5.3.37recommend
- GPT-5.1-Codex-MaxPraised for web development, debugging and multi-step refactors. Common caveat: it is slow and weak outside web development.30recommend
- MAI-Code-1-Flash by MicrosoftRecommended for strong coding benchmark scores from a small 5B-active-parameter model and a detailed technical report, though critics note weak comparisons and unpublished costs and training data.14recommend
- SWE-1.7Recommended as a fast, low-cost coding model that comes close to frontier performance on price. Most common caveat: benchmark claims are self-reported and it only works inside Cognition's own harness.14recommend
- GPT-5.1-CodexRecommended as a strong coding model that people rate above Gemini 3 Pro for usable results. Critics note it is slow, unresponsive, and sometimes misdiagnoses bugs.10recommend
- CodestralPeople recommend Codestral for coding autocomplete and generating git commit messages, and it has a free tier. Its conversation ability is weak, and some found it lackluster.8recommend
- North Mini CodeFree model people say is competitive for everyday coding tasks, but critics say it makes more mistakes and struggles with tools.6recommend
- Windsurf SWE-1.6Fast and good enough for small tasks, grunt work and tests, and people pair it with Opus for higher-level planning. Some find it slow.5recommend