Abstract:
DeepSeek V4.1 Flash has been updated to version V4.1-Flash-1017, and the official web page, APP and API models have been changed to the new model simultaneously. Compared with the previous version of DeepSeek V4.1 Flash, the new model has made great improvements in all aspects of benchmark tests.
While improving performance, we also shortened the task running time and output length of the model, allowing users to complete tasks at lower costs and faster (as shown in the figure below), achieving a new performance-cost Pareto optimal under multiple Harnesses. In particular, this time we focused on optimizing platform frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Please upgrade to DeepSeek Harness 0.2.1-rc.1 and above to get the best experience adapting to the new version of the model.
DeepSeek V4.1 Flash 1017 uses the mini-swe-v2 framework to test the performance/cost graph of DeepSWE
Complete project development from scratch
DeepSeek V4.1 Flash 1017 uses reinforcement learning (RL) training technology based on self-feedback to improve the ability to build complete projects from scratch in aspects such as game development and 3D modeling, and can support the realization of more "wish-making" programming tasks and project development requests. The following case shows the project effect directly achieved by the new version of the model through very short interaction rounds within the DeepSeek Harness framework environment with fewer prompt words and material references:
Example 1: Forbidden City 3D modeling display effect
Example 2: Development results of web racing game
Example 3: The reproduction effect of the "Portal 2" game developed based on the godot engine
Example 4: Valley of Ages-use 3D scenes to dynamically show the development process from prehistoric times to modern times
Open self-developed search API
For a long time, DeepSeek's search capabilities on the web and API have been provided with underlying support by the self-developed DeepSeek search engine, which stably handles hundreds of millions of search requests from all over the world every day.
DeepSeek search is an AI-native universal search engine that indexes hundreds of billions of web pages around the world, supports multi-language search, and returns high-quality, highly relevant, authoritative, and timely results in real time. For SEO cheating, GEO manipulation and low-quality AIGC, we filter them through strict identification mechanisms to ensure that authoritative and credible information is presented to users and LLM.
This new version released by DeepSeek API adds support for the Search and Fetch interfaces, and related capabilities are officially open to the public. The price of both interfaces is 5 yuan/thousand times.
For detailed usage instructions
Refer to the official API documentation:
API Guide - Web Search and Crawl
.
Please note that when current users use the DeepSeek API in DeepSeek Harness / Claude Code / Codex, they have already accessed the Search/Fetch interface through server-side tool calls. These Search/Fetch interface calls are currently not billed, and will start to be charged on a per-call basis two weeks later (00:00 on November 1, 2026, Beijing time).
Open source model
The download address of DeepSeek-V4.1-Flash-1017 model weights is as follows:
HuggingFace:
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
ModelScope:
https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017
Comments