DeepSeek describes its newest release in modest terms. The Hugging Face model card calls V4.1-Flash-1017 “a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” The benchmark table attached to that modest description is less modest. Released on October 17, the updated checkpoint improves on every one of the 11 benchmarks DeepSeek published, and the weights are available to anyone under the MIT license.

The biggest moves are in agent and coding tests. Terminal-Bench 4.0 rose from 31.2 to 42.3, a gain of 11.1 points. ProgramBench climbed from 20.3 to 29.1, and DeepSWE v1.1, a software engineering benchmark, rose from 74.2 to 77.0. On Humanity’s Last Exam without tools, the score went from 36.8 to 42.6. DeepSeek also says the update reduces “task execution time and output length,” which matters for cost: shorter outputs mean fewer billed tokens for the same finished job.

Abstract illustration of benchmark scores climbing across a chart
Illustration created for GPTOnline.

Strong numbers, with the source kept in view

Every figure above comes from DeepSeek’s own evaluations, published on the model card and summarized in AI Weekly’s release alert. They have not been independently replicated. DeepSeek’s own comparison table also shows where the model still trails: Claude Opus 5 leads on Humanity’s Last Exam (56.3 against 42.6) and on Terminal-Bench 4.0 (51.8 against 42.3), while the DeepSWE score of 77.0 edges ahead of Opus 5 at 74.0 in the same table. A detailed breakdown of the model card and configuration files puts the checkpoint at 763 billion parameters, with a context window of 1,048,576 tokens and text and image input.

Same name, same price, new tools

For API customers, the upgrade arrives silently: the existing deepseek-flash model name now serves the 1017 checkpoint, and reported pricing is unchanged at $0.15 per million input tokens and $0.60 per million output tokens. The release also opened two of DeepSeek’s own tools to outside developers, a Search API and a Fetch API priced at $0.75 per 1,000 successful requests. The accompanying technical report, titled “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression,” signals where much of the engineering effort went: holding a million tokens of context without the memory bill that usually comes with it.

Abstract illustration of model weights being released openly to developers
Illustration created for GPTOnline.

The open license is what turns this from a routine version bump into something the industry watches. Because the weights shipped on the same day as the API update, any lab or developer can run the checkpoint and test DeepSeek’s claims directly, rather than taking a vendor table on trust. Whether independent results match the published table is now a question of weeks, not access.

Sources