On September 8th, multiple sources confirmed that DeepSeek V4.1-Flash has officially begun its intermediate version internal testing. This version introduces a new model structure, achieving native multi-modal support while offering stronger overall capabilities, faster operation speeds, and lower calling costs.

In terms of calling methods, developers can keep the base_url unchanged and simply set the model name to deepseek-v4.1-flash-expires-on-0910 to directly call it, with the current billing standards remaining consistent with DeepSeek-V4-Flash. From external feedback so far, the most significant feature of this new Flash version is "extremely fast speed," demonstrating highly efficient pattern processing efficiency.

Looking back at its technical iteration path, multi-modal capabilities were previously a relatively weak area for DeepSeek. Just on August 21st this year, the official announced the launch of the new multi-modal visual understanding model V4-Flash-Vision-Exp on the API platform. This experimental model matches the official version in pure text capabilities (including Agent, reasoning, world knowledge, etc.), but achieved a significant improvement in the visual understanding AgentBenchmark test. Its multi-modal Agent capabilities are already very close to Opus-4.8.

In addition to the model's iteration, its open-source ecosystem has also seen rapid progress. On August 13th, DeepSeek Harness v0.1 Developer Preview was made available for global developers to test and was simultaneously open-sourced under the MIT license. After its release, the project quickly sparked discussions on overseas social platforms such as X, and within less than 24 hours, it received over 70,000 GitHub stars.

Notably, before this, DeepSeek V4 Pro had been officially launched, focusing on enhancing Agent capabilities. With the official release of the entire V4 series of models, the official also adjusted the API pricing system, introducing a peak-valley time-based pricing mechanism. The internal testing of V4.1-Flash marks another key step in its multi-modal and performance optimization journey.