DeepSeek Flash: Faster Tasks, Shorter Outputs
DeepSeek's October 17 release thread explained: agent benchmarks, project demos, DSH upgrade guidance, open weights and Search/Fetch API pricing.
On this page
What changed in V4.1-Flash-1017
On October 17, 2026, DeepSeek announced DeepSeek-V4.1-Flash-1017 in a four-post release thread. The model is available on its website, app and API. The thread also introduces Search and Fetch APIs and links to the open weights.
DeepSeek reports stronger benchmark results, faster task completion and shorter outputs than the previous version, with optimizations for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.
Official project demos
The second post describes reinforcement learning with self-feedback. The accompanying announcement shows four end-to-end projects:
- A 3D representation of the Forbidden City.
- A racing game that runs in a web browser.
- A Godot project recreating Portal 2.
- Valley of Ages, a 3D scene about different periods of history.
These demos suggest tasks to try. Start with one bounded feature and inspect the generated files, running result and retries.
Official benchmark results
These results come from the official DeepSeek model card and compare the previous DeepSeek-V4.1-Flash release with the 1017 update.
| Benchmark | DeepSeek-V4.1-Flash | DeepSeek-V4.1-Flash-1017 |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| ProgramBench | 20.3 | 29.1 |
| Automation-Bench v1.0.6 | 54.8 | 60.5 |
| Agents’ Last Exam | 31.8 | 35.4 |
These are DeepSeek's published results. Time and cost for your tasks also depend on prompts, tool configuration and retries.
Terminal-Bench 4.0 uses the Claude Code harness. DeepSWE v1.1 uses the Mini-SWE-V2 harness.
The repository and model weights use the MIT License. Weights are available on Hugging Face and ModelScope.
Upgrading DSH
The release post recommends DeepSeek Harness v0.2.1-rc.1 or later for the new model. The v0.2.1-rc.2 release notes also document support for the new model and optional Search and Fetch APIs.
Search & Fetch API
The official API guide describes two endpoints: Search API returns structured web search results, and Fetch API retrieves a page as Markdown.
POST https://api.deepseek.com/v1/tools/search
POST https://api.deepseek.com/v1/tools/fetchBoth APIs use bearer authentication with a DeepSeek API key. The Search count parameter controls the number of results and does not change the price per request.
In DSH rc.2, open Plugins → Built-in plugin settings → Web search, turn on “Use the new Web Search API” and save. The next search or fetch uses the selected APIs.
Pricing and billing dates
- Direct Search and Fetch API calls each cost $0.75 per 1,000 successful requests that return a result. Failed requests, empty search results and timeouts are not billed.
- As of October 21, 2026, Search and Fetch calls made through server-side tools in DSH, Claude Code and Codex are free for DeepSeek API users. Billing for this path starts on October 31, 2026, at 16:00 UTC, or November 1 at 00:00 Beijing time.
- The official guide checked October 21 lists independent account limits of 100 QPS for Search and 25 QPS for Fetch. Fetch also has a per-domain limit.
Read the Search and Fetch billing guide · DeepSeek’s official release thread
DeepSeek-V4.1-Flash-1017 common questions
Where can I use the new model?
DeepSeek announced availability on its website, app and API. For DSH, update Harness and check the models exposed by your configured provider.
Is DSH rc.2 required?
The post recommends rc.1 or later. The rc.2 release notes also document new-model and Search/Fetch API support, with setup steps linked above.
Are Search and Fetch free now?
Direct API calls are billed per successful request. The temporary free period for DeepSeek API users of the specified server-side tools ends November 1, 2026, at 00:00 Beijing time. Budget for the path your project actually uses.
Where are the open weights?
The fourth post links to Hugging Face and ModelScope. The model card lists an MIT license for the repository and weights.
Release thread and sources
- Official post 1/4: model release and Harness version guidance
- Official post 2/4: end-to-end project demos
- Official post 3/4: Search/Fetch and billing dates
- Official post 4/4: open weights
- Official announcement and project examples · Official model card and benchmark table · API usage and billing guide
To try DeepSeek-V4.1-Flash-1017 on your own tasks, start with DSH installation, then follow the model setup guide to connect a provider and verify a first task.