Chinese company DeepSeek has launched DeepSeek V4.1 Flash, which it presents as the smallest model in a new architecture family designed to increase inference speed and expand task-processing capabilities, with the possibility of scaling later to larger models. The model includes integrated multimodal visual understanding, enabling it to handle images alongside text-based tasks.
V4.1 Flash follows the experimental V4 Flash Vision Exp model that DeepSeek introduced in August to support visual inputs. However, the company says that multimodal capabilities have become an intrinsic part of the architecture in the new release, rather than a separate experiment within the previous model line.
Higher Performance According to DeepSeek’s Tests
Based on test results published by DeepSeek, V4.1 Flash achieved a score of 90.9 on GPQA Diamond, 90.6 on Terminal-Bench 2.1, and 74.2 on DeepSWE v1.1. The company also announced a rating of 3471 for the model on Codeforces.
DeepSeek says the model outperforms DeepSeek V4 Pro across a range of metrics, including performance, cost, speed, and the total time required to complete tasks. The material does not include detailed figures comparing prices or times, and the results cited were issued by the company itself, so they should be treated as performance claims that require independent verification before being adopted as a general benchmark.
Model Name Changes and Request Redirection
With the release of V4.1 Flash, DeepSeek will stop making V4 Flash and V4 Flash Vision Exp available. The old model names will temporarily point to V4.1 Flash to preserve compatibility with existing applications. Developers who want access to the latest Flash model through the API can use the model name deepseek-flash.
The company also plans to gradually retire V4 Pro. Starting September 14 at 12:00 Beijing time, requests sent through deepseek-v4-pro will be redirected to V4.1 Flash until V4.1 Pro is launched, with costs calculated according to V4.1 Flash pricing. The material does not state the new prices, but indicates that DeepSeek reduced API prices for the Flash series with the model’s launch.
What Changes in Practice?
For developers, the announcement is not limited to adding a new model; it also involves reorganizing access to DeepSeek models. Relying on old names may result in V4.1 Flash running instead of the original model, which is an important change when comparing results or monitoring cost and performance in production applications. In contrast, the name deepseek-flash provides a way to access the latest version in the Flash category without being tied to a specific release name.
The announcement’s structure also reveals DeepSeek’s focus on faster, lower-cost models while maintaining a level of performance aimed at the larger-model category. The company says V4.1 Flash is the smallest member of the new architecture, leaving open the possibility of launching larger models, including V4.1 Pro, within the same family. However, the timing and specifications of these releases were not specified in the material.
The launch comes at a time when Reuters reports that DeepSeek, headquartered in Hangzhou, has begun preparing for an initial public offering on the STAR Market of the Shanghai Stock Exchange. According to the source, the company may target using the new funding for computing infrastructure, model development, and retaining experienced employees. The text provides no confirmed details about the timetable or terms of the offering.