DeepSeek releases V4.1-Flash-1017 update, says it beats its predecessor on every benchmark it reported

As of 11:32 a.m. ET, Sunday, Oct. 18, 2026

Chinese AI company DeepSeek has released DeepSeek-V4.1-Flash-1017, an update to the V4.1-Flash model it launched on Sept. 10, and says the new version scores higher, finishes tasks faster and writes shorter answers than the one it replaces. The company announced the release on X at 4:35 a.m. ET Saturday, Oct. 17, saying it was "now live on our website, app, and API".

DeepSeek describes the release as "a minor update" to V4.1-Flash. The model weights and repository are published on Hugging Face under the MIT License. The company's API change log says the API calling method and pricing are unchanged and that developers reach the new model through the existing deepseek-flash model name. DeepSeek's pricing page lists DeepSeek-V4.1-Flash-1017 as the model behind that name, with a 1-million-token context length. DeepSeek's official web and app models were also switched to the new version.

What DeepSeek says the model can do

All of the benchmark figures below come from DeepSeek's own model card and change log. They are the company's results, and GROKsiNEWS.com has not seen independent testing of the model.

In DeepSeek's table, the new model scores higher than the earlier V4.1-Flash on all 11 benchmarks listed. On DeepSWE v1.1, a software-engineering benchmark, DeepSeek reports a score of 77.0 for the new model, up from 74.2 for the earlier V4.1-Flash. The same table lists OpenAI's GPT-5.6 Sol at 73.0 and Anthropic's Claude Opus 5 at 74.0 on that test. DeepSeek also reports higher scores than both of those rivals on Automation-Bench, Agents' Last Exam and Apex Bench, according to its table.

The model does not come out on top everywhere in DeepSeek's own comparison. On Terminal-Bench 4.0, DeepSeek lists the new model at 42.3 against 51.8 for Claude Opus 5. On Humanity's Last Exam without tools, it lists 42.6 against 56.3 for Opus 5 and 44.5 for GPT-5.6 Sol. Opus 5 also scores higher in DeepSeek's table on ProgramBench, NL2Repo-Bench and two of the three visual benchmarks it lists.

DeepSeek says it ran Terminal-Bench 4.0 with the Claude Code harness and DeepSWE v1.1 with the Mini-SWE-V2 harness. It recommends a context window of 1 million tokens.

The company says it also cut the model's task run time and output length, which it calls a "performance–cost Pareto" improvement. A cost chart DeepSeek released, as described by the tech outlet The Decoder, puts the new model's average cost at just under $0.20 per DeepSWE task, compared with about $0.50 for its predecessor.

The Hugging Face listing for DeepSeek's official repository shows a model size of about 763 billion parameters. That figure is calculated by the platform from the uploaded files; DeepSeek has not stated it.

Search engine opened to developers

In the same Oct. 17 change-log entry, DeepSeek said it is opening its in-house search engine to outside developers. The company says the engine already powers search in its own web app and API, handles "hundreds of millions of search requests" a day and indexes "hundreds of billions of web pages".

DeepSeek's pricing page lists the Search and Fetch APIs at $0.75 per 1,000 successful requests each. It says calls made through DeepSeek Harness, Claude Code or Codex are currently free, and that per-request billing starts at midnight Beijing time on Nov. 1, 2026, which is noon ET on Saturday, Oct. 31.

Why it matters

DeepSeek's releases are closely watched because the company publishes model weights openly and prices its API well below many U.S. rivals. Its launch post had drawn about 5.9 million views on X by early Sunday, according to the platform's metrics. The performance claims are the company's own, and how the model holds up in independent testing has not yet been established.