By use case · Coding
Best local coding models
Every model here is trained to write code, not just chat about it. 11 open coding models are ranked by quality below. The strongest is Kimi K2.7 Code (~592.1 GB at Q4_K_M); the lightest is Qwen2.5 Coder 0.5B (~1.4 GB).
Models tuned for code generation and agentic coding (SWE-bench, fill-in-the-middle).
- 1 KI ~592.1 GBKimi K2.7 Code1000B MoE · runs on 0/40 devices
- 2 ~276.2 GBQwen3-Coder 480B-A35B Instruct480B MoE · runs on 0/40 devices
- 3 ~20.7 GBQwen2.5 Coder 32B32B · runs on 13/40 devices
- 4 ~19.4 GBQwen3-Coder 30B-A3B30.5B MoE · runs on 13/40 devices
- 5 NO ~21.1 GBNorth Mini Code 1.030.5B MoE · runs on 12/40 devices
- 6 ~15.4 GBDevstral Small24B · runs on 15/40 devices
- 7 ~10.1 GBQwen2.5 Coder 14B14B · runs on 29/40 devices
- 8 ~5.8 GBQwen2.5 Coder 7B7B · runs on 33/40 devices
- 9 ~3 GBQwen2.5 Coder 3B3.09B · runs on 40/40 devices
- 10 ~2 GBQwen2.5 Coder 1.5B1.54B · runs on 40/40 devices
- 11 ~1.4 GBQwen2.5 Coder 0.5B0.494B · runs on 40/40 devices
Tagged by each model's stated purpose. The memory figure is what it needs at Q4_K_M, and the device beside it is the lightest tracked machine that fits it. "Runs on N devices" counts the 40 tracked devices that fit it at Q4_K_M.
Prefer an objective score to purpose tags? The coding leaderboard ranks these models by Aider polyglot (percent of 225 exercism exercises solved across six languages).
FAQ
What is the best local coding model?
Kimi K2.7 Code is the highest-ranked, but it needs ~592.1 GB and no tracked device runs it locally. The strongest one you can actually run is Qwen2.5 Coder 32B. Pick by what fits your memory using the list above.
What is the smallest coding model that runs on a laptop?
Qwen2.5 Coder 0.5B is the lightest, at ~1.4 GB at Q4_K_M, so any device with 8 GB or more can load it.
How were these coding models chosen?
They are open-weight models whose own design targets coding (by name, family or model card). Memory figures are computed at Q4_K_M and sourced; see the methodology page.
Memory is computed at Q4_K_M, validated 2026-08-03. See methodology.