IT之家 · 智能时代 · 10/17/2026, 16:17:20
DeepSeek Releases Open-Source V4.1 Flash-1017 with Improved Performance
DeepSeek released V4.1 Flash-1017, enhancing zero-shot project generation via self-feedback reinforcement learning. The update achieves Pareto improvements in performance and cost across benchmarks while maintaining an open-source MIT license for model weights.
SOURCE COVERAGEFull Original Report
DeepSeek V4.1 Flash-1017 Release: Enhanced Performance, Reduced Latency, and "Wish-Based" Programming Made Possible
On October 17, DeepSeek announced that DeepSeek V4.1 Flash has been updated to the V4.1-Flash-1017 version. The official web interface, app, and API models have all switched to this new model (API calling methods and pricing remain unchanged; you can still access the latest model using deepseek-flash).
According to DeepSeek's official introduction, this update achieves a Pareto improvement in performance and cost: across various benchmarks, the new version demonstrates significantly improved model capabilities compared to the previous V4.1 Flash, while simultaneously reducing task execution time and output length, thereby lowering the overall cost per call.
- HLE (w/o tools): 42.6 (44.0)
- Terminal-Bench 4.0: 42.3
- DeepSWE v1.1: 77.0
- ProgramBench: 29.1
- NL2Repo-Bench: 68.2
- Automation-Bench v1.0.6: 60.5
- Agents' Last Exam: 35.4
- Apex Bench v1.1: 72.1
- Chartography (w/tools): 83.0
- BabyVision (w/tools): 92.5
- ZeroBench-main (w/tools): 51.0
DeepSeek stated that while improving performance, the new model also reduces task execution time and output length, enabling users to complete tasks faster and at lower cost, achieving a new performance-cost Pareto optimum across multiple Harness environments.
Additionally, DeepSeek has specifically optimized compatibility for platforms such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Users are advised to upgrade DeepSeek Harness to version 0.2.1-rc.1 or higher to ensure full compatibility.
In terms of capabilities, V4.1-Flash-1017 introduces a notable change: through reinforcement learning training based on self-feedback, the model's ability to build complete projects from scratch has been enhanced, covering scenarios such as game development and 3D modeling.
Officially, this improvement enables the model to handle more "wish-based" programming requests, meaning it can directly generate runnable project outcomes even with extremely limited prompts and reference materials. Official demo cases show that within the DeepSeek Harness environment, the new version completed the construction of full projects in very few interaction rounds.
Meanwhile, the model weights for DeepSeek V4.1-Flash-1017 are now available for download. DeepSeek provides weight files on both HuggingFace and ModelScope, continuing the open-source strategy under the MIT license established by V4.1 Flash.
IT Home provided the following links: