DeepSeek releases V4.1-Flash-1017, an update it says is faster and gives shorter outputs, with open weights and a new search API

DeepSeek’s API change log, dated 17 October 2026, said DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017 and that its web and app models have been updated.^[1] The change log said the API calling method and pricing remain unchanged, and that developers reach the new version with the model name deepseek-flash.^[1] Bloomberg, in a report carried by The Economic Times, said the Hangzhou-based company unveiled the model late on Saturday afternoon in posts on X saying it had “stronger benchmark performance, faster task completion, and shorter outputs than the previous version”.^[4] DeepSeek’s model card calls it a “minor update to DeepSeek-V4.1-Flash” with improvements across reasoning, coding, agentic tasks and visual benchmarks, and says the repository and weights are licensed under the MIT License.^[2] The Decoder reported that the weights are on Hugging Face and ModelScope, and OpenRouter lists the model as released on 17 October 2026 with a 1.0 million-token context window.^[3][5]

DeepSeek’s change log and model card report scores of 77.0 on DeepSWE v1.1, 42.3 on Terminal-Bench 4.0 and 42.6 on Humanity’s Last Exam without tools (HLE), against 74.2, 31.2 and 36.8 for the previous V4.1 Flash.^[1][2] The model card’s table also lists scores for other models: on DeepSWE v1.1, 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol, while Opus 5 scores higher than the new model on several others, including HLE (56.3), Terminal-Bench 4.0 (51.8), ProgramBench (37.0) and NL2Repo-Bench (75.3), and GPT-5.6 Sol scores higher on HLE (44.5).^[2] The model card said Terminal-Bench 4.0 was run with the Claude Code harness and DeepSWE v1.1 with the Mini-SWE-V2 harness.^[2] The Decoder reported that these are DeepSeek’s own benchmarks, that the model beats larger rivals in several coding and agent categories, and that “Independent benchmarks haven’t come out yet”.^[3]

The Decoder reported that, in DeepSeek’s cost chart for DeepSWE, the model reaches 77.0% at an average cost of just under $0.20 per task, against about $0.50 for the predecessor, and that rivals in the chart cost between $2 and over $10 per task.^[3] OpenRouter lists several hosts for the model, with prices that include $0.05 per million input tokens and $0.60 per million output tokens at one host, and $0.08 and $0.50 at another.^[5] Artificial Analysis, a model benchmarking site, lists the earlier DeepSeek V4.1 Flash (Max), released in September 2026, at 39 on its Intelligence Index, ninth of 120 models, and describes it as notably fast but very verbose.^[6]

DeepSeek also opened its in-house search engine to developers through Search and Fetch APIs, and said the engine handles hundreds of millions of search requests every day.^[1] The Decoder reported that each endpoint costs $0.75 per 1,000 calls, and that DeepSeek API users can use the features free through server-side tools in DeepSeek Harness, Claude Code and Codex until 31 October 2026 (16:00 UTC).^[3]

Bloomberg reported that DeepSeek had an outage of about five hours earlier in the week that left its servers unreachable, which some experts attributed to a lack of sufficient computing resources.^[4] Bloomberg reported that the V4.1-Flash model released in early September was priced at such a steep discount that shares of MiniMax Group, Z.ai and Alibaba fell on the Hong Kong stock exchange.^[4] The Decoder reported that the Chinese government reportedly plans to tighten rules on publishing open-weight models or block their release, and that one deciding factor is whether a system ranks among the 10 most capable.^[3] The Hindu reported that on Friday 16 October 2026 China’s Ministry of Commerce and Ministry of Science and Technology announced a draft plan, “Announcement No. 48”, proposing a ban on exporting 14 technologies and restrictions on 26 more across AI, biomedicine and photovoltaics.^[7] A Ministry of Commerce spokesperson said that “most AI technologies will continue to be freely exportable” and that the draft adds 23 key AI technologies covering foundation models, intelligent agents, computing clusters and information-processing technologies.^[7] The ministry’s statement added that China remained “committed to building an open-source AI innovation ecosystem”, that open-source AI projects “would be eligible for facilitation measures”, and that technologies already public through open-source publication or academic papers would “generally be exempt from export controls”.^[7]