DeepSeek Team Member Recommends Using Minimal Mode in dsh with the Official DeepSeek V4 Pro Release
要闻
According to a blogger, a DeepSeek team member recommended using minimal mode in dsh with the official DeepSeek V4 Pro release, saying the mode is well aligned with RL training and advising users not to run it under PowerShell. The source also said that dsh's commits show the relevant configuration was added specifically to align with training.
Recently, a blogger shared leaked chat logs on social media showing that a DeepSeek team member recommended using minimalist mode in DeepSeek Harness with the official release of DeepSeek V4 Pro. In the group chat, the member said that dsh's minimalist mode is “relatively well aligned with our RL training” and encouraged everyone to try it. Regarding the runtime environment, when asked whether “linux is required,” the member explicitly replied that “at the very least, Windows PowerShell will not work.” The member added that minimalist mode is highly recommended for users seeking strong correctness and intelligence, as long as the task does not involve frontend work or other areas with aesthetic requirements.
OpenAI Updates Codex to Let the Main Agent Invoke Models Including Luna
要闻
A Codex team member introduced the Multi-agents v2 update, which allows the main Agent to delegate tasks to different supported models, including Luna. The capability had actually been added in an earlier update but received little attention at the time, prompting OpenAI staff to highlight the change more extensively.
Members of the Codex team recently highlighted the Multi-agents v2 update for Codex. The main change relaxes model restrictions for sub-agents, allowing the primary Agent to delegate work to different supported models, including Luna, based on the task. Some models can also be used as sub-agents that only execute tasks without delegating further. This change had actually been completed earlier and was subsequently included in a Codex version update. On August 15, Codex team member Eric Provencher posted that the feature had previously “not received much attention,” after which Greg Brockman and several other people involved shared the post in quick succession.
FireRed Team Releases FireRedTTS3 with Support for 21 Chinese Dialects
模型发布
小红书's FireRed Team recently open-sourced FireRedTTS3, a speech generation and editing model. It includes a Base version supporting zero-shot voice cloning across 24 languages and 21 Chinese dialects, as well as an Instruct version for voice design and editing.
Xiaohongshu's FireRed Team recently released FireRedTTS3, a unified speech generation and editing model based on semantically enhanced speech representations, as an open-source project under the Apache-2.0 license. It includes two versions: FireRedTTS3-Base, released first, supports zero-shot voice cloning across 24 languages and 21 Chinese dialects; FireRedTTS3-Instruct was released later with the model and PyTorch code, supporting natural-language-based voice design as well as both semantic and acoustic speech editing. According to official data, FireRedTTS3-Base achieved an average WER of 3.04% and speaker similarity of 78.8% on Seed-TTS-eval, as well as an average WER of 3.754% and similarity of 84.8% on MiniMax-MLS-Test. The team says these are the best results among the models compared.
Devin Offers SWE-1.7 and GLM-5.2 Free for One Month to Paid Subscribers
开发生态
Devin recently announced that SWE-1.7 and GLM-5.2 will be available free for one month to subscribers on plans including Pro, Max, and Teams. Both models are available immediately in Devin Desktop and Devin CLI.
Devin recently announced that SWE-1.7 and GLM-5.2 will be available free of charge for the next month to all Pro, Max, and Teams subscribers. Users can continue using both models through Devin Desktop and Devin CLI. The free-access period begins immediately and will last for a full month.
Note: This content was created with AI assistance and may contain hallucinations and errors.
