DeepSeek ships V4.1-Flash-1017 update with open weights under MIT licence

What happened

DeepSeek reports gains in reasoning, coding, agent tasks and visual benchmarks, along with shorter outputs.

The largest vendor-reported change is on Terminal-Bench 4.0, where the score rose from 31.2 to 42.3. The API price is unchanged.

The architecture is the same as V4.1-Flash. No independent replication of the new scores was available at release, so the gains rest on DeepSeek's own figures for now.

Weights are on Hugging Face under the MIT licence. The model is also live on DeepSeek's website, app and API.

Why it matters for creators

Cheap open-weight models keep lowering the cost of prompt-writing and captioning steps in creator workflows.

Shorter outputs can also mean fewer tokens per request for the same job, and the open weights give teams the option to run the model on their own hardware as well as through DeepSeek's API.

As with any vendor benchmark, the reported gains are worth testing on real tasks before switching a pipeline over.