ChatGPT Work launches a feature that learns users' personal writing styles
要闻
OpenAI announced that personal writing-style learning is now available to every paid subscription plan with access to ChatGPT Work. Once users connect their everyday tools, the system learns their expression patterns from emails, messages, and files and applies them to future writing.
OpenAI says personal writing-style learning is currently available on the web to all paid subscription plans that include access to ChatGPT Work. Users can connect everyday tools such as Gmail, Google Drive, Slack, and SharePoint to ChatGPT Work, allowing the system to learn preferred phrases, specific sign-off formats, and capitalization habits from their emails, messages, and files and carry those patterns into later writing. Configuration is available through the official web settings entry or by navigating to Personalization and then Writing Style in Settings; once configured, the style applies to all messages in ChatGPT Work on both web and mobile.
Codex head Tibo announces usage resets for all paid subscriptions
要闻
Codex lead Tibo announced a one-time usage reset for all paid subscriptions at around 18:00 Pacific Standard Time. According to his explanation, the measure is intended to let users continue using Astra after spending usage on 3D modeling in Blender.
Codex lead Tibo said a usage reset covering all paid subscriptions is expected to take effect at around 18:00 Pacific Standard Time. The measure does not distinguish among paid subscription types and applies to every paid subscription. According to his explanation, the reset is intended to let users continue using Astra even after first consuming usage on 3D modeling in Blender; he also mentioned that the workweek was about to begin. The specific implementation time is based on the approximately 18:00 PST timing he provided.
腾讯混元 fully rolls out dedicated optimizations for Hy4 preview
要闻
The joint Tencent Hunyuan and WorkBuddy team has fully rolled out an optimized Hy4 preview release addressing prolonged reasoning and excessive self-verification in complex tasks. The team says it reduced task rounds and input and output token usage without compromising task outcomes.
On September 8, 2026, the joint Tencent Hunyuan and WorkBuddy team announced the full rollout of an optimized release for Hunyuan Hy4 preview. The model was previously launched first on WorkBuddy, and the team says it subsequently received usage feedback from users and developers. The changes target prolonged reasoning and excessive self-verification during complex tasks, with the team monitoring the results through two methods: Bench metrics and human evaluation. According to its explanation, the new release reduced task rounds and input and output token usage without compromising task outcomes. The team said it would continue iterating and collecting user feedback and suggestions.
OpenBMB has open-sourced MiniCPM5-2B, a 2B-parameter language model for on-device use, local deployment, and resource-constrained settings. According to OpenBMB, it averaged 53.9 across 34 benchmarks and scored 23 on the Artificial Analysis intelligence index.
OpenBMB has officially released MiniCPM5-2B, the second model in the MiniCPM5 series, and made its model weights and code available under the Apache-2.0 license. Following MiniCPM5-1B, the 2B-parameter language model targets on-device use, local deployment, and resource-constrained settings. According to OpenBMB, MiniCPM5-2B reached SOTA among open-source models of the same size, averaged 53.9 across 34 benchmarks, and outscored every larger model included in the comparison. It also scored 23 on the Artificial Analysis intelligence index, ranking first among open-source models with 4B parameters or fewer.
MiniCPM5-2B uses the standard LlamaForCausalLM architecture and has a context length of 131072. OpenBMB also released the UltraData training datasets, training recipes, and RL technology stack. Day-0 adaptation has been completed for Intel, Arm, and Rockchip platforms, while GGUF, MLX, GPTQ, and other formats are available. The model can be obtained through Hugging Face, GitHub, and ModelScope.
TRAE has launched a support program for current university students and teachers across China, excluding Hong Kong, Macao, and Taiwan, running through December 31, 2026. Students in participating classes can receive up to 8,000 points each, while teacher participation is capped at 100 places.
TRAE has launched a university support program for current students and teachers across China, excluding Hong Kong, Macao, and Taiwan, with the program running through December 31, 2026. Current students can receive 2,000 points after completing student verification, while new users receive an additional 4,000-point welcome bonus and can claim 500 points on their first login each month. Each allocation of points remains valid for 90 days from its issuance date. When a university teacher enrolls in the TRAE 先锋教师计划 and uses TRAE to support a course, every student in the class receives an additional 6,000 points, bringing the total with student verification points to a maximum of 8,000 per student. The program is limited to 100 teachers, with places allocated in registration order. Students can complete verification through the mobile “TRAE 大学生专属福利页” or the “学生认证” entry under account management on desktop, while teachers can apply through “合作招募” in the menu of TRAE’s official WeChat account.
Tencent Hy Team has open-sourced the MoE model Hy4 preview, with 770B total parameters and 49B activated per token; it averaged 2.99 in an internal blind evaluation covering 203 engineering tasks.
Tencent Hy Team has released Hy4 preview under the Apache License 2.0. The model has 770B total parameters, with 49B activated per token. Its 78-layer backbone uses a dense FFN in the first layer and MoE in the remaining 77 layers. Each MoE layer contains 256 routed experts and 1 shared expert, with every token activating the top-8 routed experts and the shared expert. The model also includes 1 native MTP layer for speculative decoding, with 10B total parameters and 0.7B activated. Its attention module uses Gated DeepSeek Sparse Attention with IndexCache to reuse sparse indexes across layers, while the residual pathway uses iHC to expand inter-layer information flow.
According to the team, Tencent specialists in software engineering, game development, financial analysis, security, and other fields participated in building the training data, while the model was co-designed with CodeBuddy and WorkBuddy. In a blind evaluation, 163 internal experts assessed outputs on 203 engineering tasks. Hy4 preview scored an average of 2.99 versus 2.92 for GLM 5.3, with win, tie, and loss rates of 46.8%, 12.8%, and 40.4%. Against Kimi K3, the scores were 2.99 versus 2.94, with rates of 51.2%, 7.9%, and 40.9%. The team says known issues in the current version include spending longer than necessary reasoning through complex tasks and a tendency to over-verify its own work. The model can be deployed with vLLM or SGLang and called through an OpenAI-compatible API. The recommended parameters are temperature=0.9 and top_p=1.0, while reasoning mode defaults to high. Direct responses can be requested with extra_body={"chat_template_kwargs":{"reasoning_effort":"no_think"}}. The project also provides a complete finetuning pipeline and the AngelSlim compression toolkit.
Qoder has launched its digital employee product QoderWake 1.0, with Mac and Windows downloads and cloud-host deployment now available. Its preset roles have increased from 6 in v0.1 to 10, alongside group collaboration, workflow reuse, and a unified task board.
Qoder has officially released QoderWake 1.0, making the Mac and Windows versions available for download and supporting deployment to cloud hosts. The release expands its preset specialized roles from 6 in the v0.1 public beta to 10, adding a project manager, UI designer, DevOps engineer, and group-chat support specialist called Q仔. Each digital employee is called a Waker and can remain active in DingTalk, Feishu, and WeCom groups, receive tasks through @ mentions, or be triggered by schedules, keywords, and events for 7×24 operation. New features include Waker groups, collaboration SOPs, WakerFlow for workflow reuse, and a unified task board. According to Qoder, sensitive data is stored locally, high-risk operations are blocked by default, and organization-level permissions can be used to configure Wakers as shared digital employees for teams.
千问开放平台 adds more than 10 AI agents for financial services
产品应用
Qwen Open Platform has added more than 10 financial-service Agents. Users can access AI services spanning securities investment, funds, futures, and insurance through one entry point, then start a conversation by @-mentioning a service or using the home-screen badge.
Qwen Open Platform has launched more than 10 financial-service Agents covering securities investment, funds, futures, insurance, and other fields. Users can @-mention a relevant service in Qwen or click the “round badge” in the upper-right corner of the home screen to enter the corresponding Agent. Available services include Industrial Securities’ intelligent investment assistant, Jiufang Lingxi Investment Assistant, and ZhongAn Insurance. According to the platform, these Agents can organize trending A-share sectors and the related industry, policy, or capital flows; analyze long-term investments in technology stocks with reference to Warren Buffett’s investment strategy; and compare underwriting requirements against a user’s health status to identify insurance products available under standard terms.
DeepSeek is hiring senior backend and server-side engineers
行业动态
DeepSeek has opened about 150 positions, primarily for senior backend and server-side engineers with 2–10 years of experience, while also accepting recent graduates and emerging talent. The roles are based in Beijing or Hangzhou, with Beijing preferred.
DeepSeek has opened about 150 positions, primarily recruiting senior backend and server-side engineers with 2–10 years of experience while also accepting recent graduates and emerging talent. The openings include server-side development engineers and Agent elastic computing R&D engineers, covering large-model research platforms, Agent framework components, R&D productivity infrastructure, the DeepSeek API, online services, data engineering, platform development and maintenance, and low-level optimization work. The positions are based in Beijing or Hangzhou, with Beijing preferred. While sharing the recruitment information, a member of the official staff said that, in their personal assessment, growth in the scale of data, machines and containers, training and evaluation tasks, Agent environments, and user requests is increasing the complexity of backend systems, creating a need to hire more engineers to upgrade, maintain, and rewrite the related systems.
ChatGPT regains website traffic share as Gemini falls back to 25.6%
行业动态
Similarweb’s AI chatbot website traffic data shows ChatGPT’s share rose from 52.7% to 55.5% over three months, while Gemini fell from 27.8% to 25.6%.
In Similarweb’s measurement of AI chatbot website visits, ChatGPT currently ranks first at 55.5%, compared with 52.7% three months earlier. Gemini now accounts for 25.6%, down from 27.8% three months earlier. Year over year, ChatGPT fell from 73.3% to 55.5%, Gemini doubled its share, and Claude rose from 1.9% to 9.3%. The current shares of the remaining platforms are 3.4% for DeepSeek, 2.4% for Grok, 1.6% for Copilot, and 0.9% for Perplexity. These figures cover website traffic only. The article says Google may use its Android ecosystem to open the Gemini app directly through push notifications and show a summary instead of a web link, leaving the related mobile interactions outside the data. OpenAI’s mobile ChatGPT app and desktop version ChatGPT Work are also excluded.
