DeepSeek’s V4.1 Flash update posts broad benchmark gains, opens Search and Fetch APIs

DeepSeek released V4.1-Flash-1017 on October 17, describing it as a minor update to V4.1 Flash with gains in reasoning, coding, agentic work and visual benchmarks. The company says the model completes tasks faster and produces shorter outputs. It is available through DeepSeek’s website, app and API, and its weights are published on Hugging Face and ModelScope.

DeepSeek’s model card reports higher scores than the previous Flash version in all 11 benchmark rows it lists. Terminal-Bench 4.0 rose from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 from 74.2 to 77.0. DeepSeek says it used Claude Code for Terminal-Bench and the Mini-SWE-V2 harness for DeepSWE; these are company-reported results, not independent evaluations.

Comparisons with other models in the same table are mixed. V4.1-Flash-1017 scores above the listed Opus 5 results on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench, but below on the other seven listed tests, including HLE, Terminal-Bench and the three visual benchmarks. Those visual results are marked as using tools, so they reflect a model-and-tools setup rather than the model alone.

An independent benchmark aggregator, BenchLM, lists the model at 65.6 out of 100 overall, ranked 35th among 224 models in its catalog. Its category ranks place the model 18th of 150 in coding and 24th of 128 in agentic capability. BenchLM also flags uneven coverage: its page shows 33 benchmark rows with displayable evidence against 681 tracked slots, and says the model’s lowest eligible category is knowledge. These rankings are specific to BenchLM’s scoring and comparison system, not a universal measure of quality.

DeepSeek’s launch material highlights examples of end-to-end project generation, including a web racing game, a Godot recreation of Portal 2 and 3D scenes. The company attributes the capability to reinforcement learning with self-feedback. The selected examples show what DeepSeek says the model can produce, but they do not establish a general success rate for building projects.

The model card specifies a 1-million-token context window, provides instructions for local inference and lists the weights under the MIT license. It also includes prompt-encoding reference code and a Rust toolkit with Python bindings. The toolkit handles conversion and parsing of model requests and outputs; DeepSeek says inference, tool execution and HTTP transport remain the caller’s responsibility.

DeepSeek also opened its Search and Fetch APIs to all users, pricing each endpoint at $0.75 per 1,000 calls. The company says calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are currently free for DeepSeek API users, with billing scheduled to begin October 31, 2026, at 16:00 UTC. Its claims of faster tasks and shorter outputs do not by themselves show that the update will cost less for every workload; costs will depend on usage and the tools involved.