DeepSeek-V4.1-Flash advances in coding, reasoning, and visual understanding
DeepSeek updates V4.1-Flash with gains in coding, reasoning, and vision, and opens its Search & Fetch APIs to developers.
Building a video game, developing an application, or modeling a 3D scene from written instructions. With DeepSeek-V4.1-Flash-1017, DeepSeek continues refining its model for agentic software development. Released on October 17, 2026, the update improves on its predecessor across reasoning, coding, automation, and visual understanding benchmarks. The company also reports shorter task execution times and more concise outputs, without changing API pricing. Alongside the model, DeepSeek is opening its Search and Fetch APIs to developers, making its in-house search infrastructure available beyond its own applications. An Update Focused on Completing Tasks
DeepSeek-V4.1-Flash-1017 succeeds DeepSeek-V4.1-Flash, introduced in September 2026.
It retains the family's multimodal capabilities, allowing it to process text and images, while strengthening its performance in software development workflows.
Coding agents do more than write code. They need to understand requests, inspect files, use tools, execute commands, and revise their work when initial attempts fail.
The updated model targets this sequence of operations.
DeepSeek reports using reinforcement learning with self-feedback to improve performance on multi-step development tasks.
The company demonstrates projects created from initial instructions, including games, interactive environments, and three-dimensional modeling.
These demonstrations illustrate the intended capabilities, although they do not establish how reliably the model completes unfamiliar projects. Stronger Software Engineering Benchmarks
DeepSeek reports improvements across several technical evaluations.
On Terminal-Bench 4.0, which assesses tasks performed through a terminal environment, the model improves from 31.2 to 42.3.
On ProgramBench, the score increases from 20.3 to 29.1.
DeepSWE v1.1, which evaluates software engineering problem-solving, rises from 74.2 to 77.0.
On NL2Repo-Bench, covering changes to code repositories based on natural-language instructions, the model scores 68.2 compared with 65.4 previously.
Automation results also improve.
On Automation-Bench v1.0.6, DeepSeek-V4.1-Flash-1017 reaches 60.5, up from 54.8.
These measurements reflect tasks in specific benchmark environments. They should not be interpreted as general percentages of software projects completed successfully in production. Improvements in Reasoning and Agent Tasks
DeepSeek also reports gains on benchmarks extending beyond conventional coding.
On Humanity's Last Exam without tools, the model scores 42.6 compared with 36.8 for the previous release.
On Agents' Last Exam, performance rises from 31.8 to 35.4.
Apex Bench v1.1 improves from 64.6 to 72.1.
Some of the new scores exceed those reported for more expensive general-purpose models.
On Automation-Bench, for example, DeepSeek-V4.1-Flash-1017 reaches 60.5, compared with 45.8 for GPT-5.6 Sol and 50.3 for Claude Opus 5 in DeepSeek's published comparison.
On ProgramBench, it scores 29.1, ahead of GPT-5.6 Sol's 23.0 but behind Claude Opus 5's 37.0.
The relative rankings therefore depend on the task being evaluated.
These are benchmark results reported by DeepSeek, not an independently reproduced evaluation conducted under identical conditions for every provider. Better Visual Understanding
DeepSeek-V4.1-Flash-1017 retains native image understanding without requiring a separate vision model.
This capability is useful for agents inspecting interfaces, interpreting charts, or extracting information from screenshots.
On Chartography, with tools enabled, the score increases from 78.9 to 83.0.
On BabyVision, it reaches 92.5 compared with 89.6 previously.
ZeroBench-main shows a smaller improvement, moving from 49.0 to 51.0.
These results suggest the update extends beyond text reasoning and coding.
Visual understanding can be important when an agent needs to inspect the outcome of an action before deciding what to do next.
It may also help agents operate graphical applications in which important information is only available through the user interface. Shorter Outputs to Reduce Cost per Task
Another change highlighted by DeepSeek concerns execution efficiency.
The company says the updated model completes tasks faster and produces shorter responses than its predecessor.
Reducing output length can lower token consumption, particularly for agents making repeated model calls.
A coding assistant may generate many intermediate responses while reviewing a project, editing files, and validating results.
Reducing unnecessary output while maintaining decision quality can lower the cost of the overall workflow.
DeepSeek describes this as an improvement along the performance-cost Pareto frontier.
However, the release does not introduce another reduction in the per-token API rate. The claimed savings primarily come from improved task efficiency and fewer generated tokens.
The company has not published a single percentage that would represent savings across every use case. Optimized for Agent Development Environments
The updated model is optimized for several agent environments, including DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.
These systems provide the tools required for software development, including file access, code editing, command execution, and interaction with external services.
DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the updated model.
Performance may vary depending on the agent framework and configuration.
For Terminal-Bench 4.0, DeepSeek used the Claude Code harness.
For DeepSWE v1.1, it used Mini-SWE-V2.
This distinction matters because an agent's performance depends on both the underlying model and the tools through which it acts.
Changing the execution environment can affect results even when the model remains the same. Search and Fetch APIs Open to All Developers
Alongside the model update, DeepSeek is making two web search APIs publicly accessible.
Search lets developers query DeepSeek's in-house search engine for results relevant to a