Best open-source models for coding

The best open-source or open-weight coding models are ranked here by SWE-bench score per effective dollar, then compared by price, context, and host availability. The current classification is provisional: verify each model's licence before commercial use.

#ModelSWE-benchCheapest hostInput $/MTokOutput $/MTokHostsContext
1DeepSeek V4 Flash 073188.8%Relace$0.04$0.08321.048576M
2DeepSeek V4 Pro 081396.4%Alibaba$0.58$1.74151M
3Qwen3.6 Plus73.4%Alibaba$0.33$1.9511M
4Qwen3.6 27B70.0%Chutes$0.29$2.0013262k
5DeepSeek V4 Pro 042377.4%StreamLake$0.57$1.15231.024M
6Qwen3.8 27B86.0%Alibaba$0.35$2.55101M
7Kimi K2.570.0%ModelRun$0.40$1.9020262k
8Kimi K2.676.2%Decart$0.55$2.3128262k
9Nemotron 3 Ultra69.0%DeepInfra$0.50$2.206262k
10Devstral 2 251262.8%Mistral$0.44$2.201262k
11Kimi K2 071160.2%Novita$0.57$2.302131k
12Kimi K2.7 Code78.2%Inceptron$0.67$3.4021262k
13Qwen3.6 Max Preview72.8%Alibaba$1.03$6.161262k
14Qwen3.7 Max68.8%Alibaba$1.48$4.4211M
15Qwen3.8 Max85.6%Alibaba$2.00$6.0011M
16Kimi K393.4%Sail Research$2.60$13.0015975k

How this ranking works

Models must have a benchmark score at or above the site's coding threshold. The ranking uses SWE-bench divided by effective blended token cost. Host prices are compared separately; a cheaper host does not imply that the model itself is licensed for commercial use.

Open-source and open-weight are not interchangeable legal categories. Check the model licence and the provider's hosting terms before using any model in a commercial product.