DeepSeek upgrades Flash for coding agents and opens its web search tools

DeepSeek upgraded its Flash model across its website, app and API on October 17, 2026, while publishing downloadable weights and opening its web search tools to developers. The new V4.1-Flash-1017 promises better coding and multi-step task performance without changing API prices. DeepSeek says the savings come from faster runs and shorter outputs.

A new snapshot behind the existing endpoint

The release follows the original V4.1-Flash launch on September 10. That version introduced native visual understanding to the model family. DeepSeek describes the October release as a minor update, rather than a new model generation, but says it improves reasoning, coding, visual benchmarks and tasks where an AI agent works through several steps.

Existing API customers do not need a new model name: deepseek-flash now accesses the updated version, with the calling method unchanged. The official website and app have also switched. For teams that want the model files, DeepSeek has published the weights on Hugging Face and ModelScope; the Hugging Face repository and weights carry the MIT license.

The model card lists a 1-million-token context window—the amount of material the model can work with at once. It also recommends setting max_tokens to at least 256,000. Those are separate settings: one describes the working context, while the other controls the allowed generation length.

Higher scores, shorter runs

DeepSeek’s benchmark table shows improvements over the previous Flash snapshot across its listed tests. Two coding results illustrate the difference: Terminal-Bench 4.0 rose more sharply than DeepSWE v1.1. The company tested the former using Claude Code and the latter using Mini-SWE-V2, the software frameworks that organize the agent’s work during evaluation.

DeepSeek’s reported coding scores

Terminal-Bench 4.0

31.2→42.3

score

Previous V4.1-Flash to V4.1-Flash-1017, according to DeepSeek’s model card.

DeepSWE v1.1

74.2→77.0

score

Previous V4.1-Flash to V4.1-Flash-1017, according to DeepSeek’s model card.

The cost claim concerns completed work, not a fresh reduction in token prices. In DeepSeek’s DeepSWE cost chart, described by The Decoder, the new model scored 77.0% at just under $0.20 per task. Its predecessor scored 74.2% at around $0.50 per run. A lower setting cleared 72% for about $0.10. These are company results; independent benchmarks were not yet available when The Decoder published.

For complete software projects, DeepSeek says it used reinforcement learning with self-feedback. The model’s own confidence scores or self-generated critiques provide the training reward signal, rather than human evaluators or an external reward model. The company links that method to end-to-end development tasks, including games, 3D modeling and design work.

DeepSeek also says it optimized the update for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. It recommends upgrading its own Harness software to version 0.2.1-rc.1 or later. That recommendation sits alongside the unchanged API access: selecting the latest model and updating the surrounding agent software are distinct parts of adopting the release.

Web access opens; deployment still needs its own plumbing

Alongside the model update, DeepSeek opened access to its in-house search engine through Search and Fetch APIs. Each endpoint costs $0.75 per 1,000 calls. The company says the underlying engine already handles hundreds of millions of search requests daily and indexes hundreds of billions of web pages. Those scale figures are DeepSeek’s own description of the service.

There is a temporary free route for existing API users: server-side Search and Fetch tools in DeepSeek Harness, Claude Code and Codex remain free until October 31, 2026, at 4:00 PM UTC. That offer is tied to those tool integrations, rather than a statement that every use of the new endpoints is free.

For developers using the downloadable model, the release also supplies prompt-handling tools. It does not include a Jinja-format chat template. Instead, the repository provides a Python reference implementation and deepseek-recipe, a maintained toolkit built from Rust libraries with Python bindings. The two options cover how conversations and model output are encoded and parsed.

• The Python reference includes tests for multi-turn conversations, tool calling, thinking mode, reasoning effort, system messages within a conversation and interleaved images.
• deepseek-recipe converts Messages, Chat Completions and Responses requests into DeepSeek’s conversation format, then parses complete or streamed model responses. Running the model, executing tools and handling HTTP connections remain the caller’s responsibility.