DeepSeek's smallest model just got faster, cheaper and better at coding.

DeepSeek released V4.1-Flash-1017 with open weights under an MIT license. In DeepSeek's own tests it scored 77% on the DeepSWE coding benchmark at just under 20 cents per task, ahead of bigger models from Anthropic and OpenAI. API prices aren't changing, and DeepSeek is also opening its in-house search engine to developers.


