DeepSeek-V4.1-Flash-1017 Released
DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated version of its Flash model. The company says it improves coding, agentic tasks, speed and output efficiency, while using shorter responses to finish work faster.
For developers who run AI agents, this matters because cost per task adds up quickly. Here is what has changed and what you should treat carefully.
Quick Look at the Update
- Model: DeepSeek-V4.1-Flash-1017
- Focus: coding, agentic tasks, speed and efficiency
- Availability: DeepSeek’s website, app and API
- Open weights: published on Hugging Face and ModelScope
- New for developers: Search and Fetch APIs
What Is New in This Model?
DeepSeek says the update beats its predecessor on benchmarks while completing tasks faster and writing shorter responses. The company describes this as a Pareto improvement, meaning performance and cost both move in a better direction instead of trading one for the other.
A New Training Method: RLSF
For end-to-end software development, DeepSeek uses a post-training method called reinforcement learning with self-feedback (RLSF).
In simple terms, the model uses its own confidence scores or generated critiques as reward signals. It does not rely on human reviewers or separate external reward models.
DeepSeek says this improves the model’s ability to build complete projects from scratch, including:
- Game development
- 3D modeling
- Design work
- Compact animated explainer videos
Built for Specific Development Tools
DeepSeek tuned the model for environments including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. For DeepSeek Harness, the company recommends version v0.2.1-rc.1 or newer.
How It Performs: DeepSeek’s Own Benchmarks
According to DeepSeek, V4.1-Flash-1017 improves over the previous V4.1-Flash in every tested category. The company also claims this smaller model outperforms larger systems, including Claude Opus 5, GPT-5.6 Sol, Kimi K3 and GLM 5.3, in several coding and AI agent benchmarks.
The clearest numbers come from the DeepSWE benchmark:

DeepSeek’s chart also places models such as GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 at roughly $2 to more than $10 per task, with lower accuracy than the new Flash model.
An Important Caution
These results come from DeepSeek’s own testing. The report states that independent benchmarks are not yet available, and that real-world developer opinions about code quality may differ from benchmark scores.
If you plan to use it for real projects, test it on your own codebase first. A model that scores well on a benchmark does not always write code that fits your project’s style, tools and edge cases.
DeepSeek Opens Its Search Engine to Developers
Alongside the model, DeepSeek is opening its internal web search technology to outside developers for the first time. The new Search and Fetch APIs use DeepSeek’s own search engine, which the company says handles hundreds of millions of queries per day worldwide.
- Price: $0.75 per 1,000 calls for each endpoint
- Free access: DeepSeek API users can use the tools for free through server-side integrations in DeepSeek Harness, Claude Code and Codex until October 31, 2026, at 4:00 PM UTC
Possible Limits on Open-Weight Models in China
The release comes as the Chinese government is reportedly considering tighter controls on publishing open-weight AI models and other key AI technologies. Under the reported proposal, whether a model ranks among the 10 most capable systems could be one factor in deciding if extra restrictions apply.
This is reported, not confirmed. If such rules do arrive, they could make it harder for developers to depend on Chinese open-weight models in the long term.
Who Should Pay Attention?
It is worth a look if you:
- Run coding agents and care about cost per task
- Want open weights you can download and test
- Build tools that need web search and page fetching through an API
Be careful if you:
- Need independently verified results before choosing a model
- Plan long-term projects that depend on open-weight availability
Quick Summary
- DeepSeek released V4.1-Flash-1017 with improvements in coding, agent tasks, speed and efficiency
- It is available on the website, app and API, with open weights on Hugging Face and ModelScope
- RLSF lets the model use its own feedback as a reward signal
- DeepSeek claims 77.0% on DeepSWE at under $0.20 per task, versus 74.2% at about $0.50 for the earlier model
- All benchmark claims are the company’s own, and independent tests are not yet available
- New Search and Fetch APIs cost $0.75 per 1,000 calls, with free access in some tools until October 31, 2026
