Key Info

DeepSeek has released V4.1-Flash-1017, a model that beats GPT 5.6 Sol on Terminal-Bench while being roughly 17x cheaper. The weights are open sourced, and it can be tried in Cline via the CLI or desktop app.

Highlights

  • Beats GPT 5.6 Sol on Terminal-Bench at about 17x lower cost
  • Improves on the previous DeepSeek V4.1-Flash result by 35%
  • Gain achieved with "no algorithmic innovation," instead via RL training on multiple coding agent harnesses
  • Model weights are already open sourced
  • Available to try in Cline via npm i -g cline or the desktop app