DeepSeek now runs its web, app and API services on a new model, V4.1-Flash-1017, Pandaily reported on 18 October. The news came from DeepSeek's announcement, relayed by IT Home. Developers still call deepseek-flash, and DeepSeek says the price is unchanged. Its model card shows Terminal-Bench 4.0 climbing from 31.2 to 42.3. Hugging Face and ModelScope host the open weights under an MIT license. Vercel said on 17 October that its AI Gateway lists the dated name deepseek/deepseek-v4.1-flash-1017.
DeepSeek says it trained the model with self-feedback, meaning the model critiques its own finished work during training. The company credits this for building a runnable project from a short prompt. DeepSeek also says the update runs faster and writes shorter outputs, which lowers the cost per call. These are the company's claims. The scores come from its own model card, and the sources show no outside tests.
An unchanged API name does not mean an unchanged model. Teams on deepseek-flash now get different behavior without making a change. Re-run your own tests on real tasks. Decide whether to pin the dated snapshot or self-host the MIT weights.