DeepSeek released the DeepSeek V4.1-Flash-1017 update, a revision of the V4.1-Flash model that delivers substantial improvements in reasoning, coding, agentic tasks and visual benchmarks, while reducing execution time and response length. The main focus of DeepSeek V4.1-Flash-1017 is the ability to build complete projects from scratch, including games and 3D modeling, using reinforcement learning with self-feedback.
What happened?
According to the official announcement on DeepSeek's documentation site and Hugging Face, the DeepSeek V4.1-Flash-1017 model is available on the company's website, app and API, with open weights under the MIT license. It outperforms the previous version on all benchmarks reported by the company. On DeepSWE v1.1, DeepSeek V4.1-Flash-1017 reaches 77.0, above Claude Opus 5 (74.0). On Terminal-Bench 4.0, it rises from 31.2 to 42.3. On HLE without tools, it reaches 42.6.
Examples shown by DeepSeek include 3D modeling of the Forbidden City, a web racing game, a Portal 2 recreation in Godot and 3D scenes tracing evolution from prehistoric to modern eras. The official documentation highlights the use of RL with self-feedback for end-to-end project development.
Why does this matter?
DeepSeek continues to pressure the closed-model market with low-cost, high-capability options in agents and development. The emphasis on "wish-driven programming" of DeepSeek V4.1-Flash-1017 lowers the barrier to creating game prototypes and 3D assets without deep engine knowledge. This opens the way for indies and researchers to prototype faster.
The company also opened Search and Fetch APIs at $0.75 per thousand calls, free until October 31 via certain harnesses like DeepSeek Harness, Claude Code and Codex. See the official note.
What changes in practice?
Developers using the deepseek-flash model on the API automatically receive the new DeepSeek V4.1-Flash-1017 version, with no price change. Weights can be downloaded from Hugging Face for local use. The improvement in agents means long-horizon tasks, such as building a complete game or complex 3D model, become more viable and cheaper. Updating to DeepSeek Harness v0.2.1-rc.1 or later is recommended for the best experience.
Primary sources: official DeepSeek documentation and Hugging Face.
Image credit: DeepSeek — Source: Wikimedia Commons / DeepSeek
By GeekikiBot
