DeepSeek-V4-Flash official version is online for public testing with super smart performance GLM-5

The official version of DeepSeek-V4-Flash is launched for public testing. The official version of GLM-5.2 DeepSeek-V4-Flash (V4 Flash) with ultra-smart performance is launched for public testing today. According to the official introduction, the capabilities of V4 Flash Agent have been greatly enhanced, and in multiple benchmark tests, its performance far exceeded that of V4-Pro-Preview, GLM-5.2 and other models. For the Code Agent task in the public benchmark test set, the official version of V4 Flash uses the minimalist mode of DeepSeek Harness (soon to be released) as the framework for testing, and uses the max gear, topp=0.95, temperature=1.0. At present, the official version of V4-Flash has natively supported the Responses API format and has been optimized and adapted to Codex. The model structure and size of V4-Flash-0731 remain the same as V4-Flash-Preview, and only the post-training is re-trained. This time, only the API interface of DeepSeek-V4-Flash has been upgraded. The DeepSeek-V4-Pro API and APP/WEB side model have not been changed. The official version of DeepSeek-V4-Pro will be released as soon as possible. via cnBeta.COM - Chinese industry information station (author: Source: Sina Technology)