Alibaba uses proprietary T-Head chips to train new Qwen model
Read more
TechNode
technode.com

Alibaba uses proprietary T-Head chips to train new Qwen model

CEO Eddie Wu stated during the company's quarterly report that Alibaba is utilizing chips developed by its semiconductor division, T-Head, to train the next version of the Qwen model.

Furthermore, Alibaba's Bailian platform provides inference services using Chinese-made chips. This expands T-Head's scope across both key stages of AI operation: model training and deployment for responding to user queries.

Bailian, known internationally as Model Studio, offers access to tools and models for building artificial intelligence applications. According to Wu, the integration of Alibaba's proprietary chips and models will reduce computing costs and increase application development efficiency.

He also emphasized that the company's complete AI infrastructure will give it an advantage within China's internal supply chain of computing power. The training process links two fundamental areas of Alibaba's AI investment: the Qwen model family and the hardware supporting cloud services.

Similar stories

Alibaba's QwenWork adds third-party models from Kimi, GLM, DeepSeek, and others
Read more
pandaily.com

Alibaba's QwenWork adds third-party models from Kimi, GLM, DeepSeek, and others

Alibaba's office agent QwenWork expanded its model selection menu on November 16th, including third-party options from Kimi, GLM, DeepSeek, MiniMax, and Step, in addition to Qwen's proprietary models. According to a report by Leiphone, this selection is available to users in both standard and enterprise versions.

This change allows users to choose the model that will perform an office task instead of relying solely on Alibaba's existing options: Standard, Advanced, and Qwen3.8-Max. Alibaba explained that this expansion is due to differences in reasoning capability, response speed, and token consumption, as users assign increasingly specialized tasks to the agents.

The published model menu includes Kimi-K3.1-Pro, GLM-5.3, GLM-5.3-Flash, DeepSeek-V4-Pro, DeepSeek-V4.1-Flash, MiniMax-M3.1-Flash, and Step-5-Preview. The recently released Qwen4-Max from Alibaba is also available. These options are part of QwenWork, and the update does not represent a joint release of models by the companies listed in the menu.

Alibaba recommends Qwen4-Max for tasks requiring multi-step complex reasoning. The company claims that this model can provide an initial working version of the result and then expand and refine it based on feedback. Furthermore, the company describes planning, decision-making, and execution regarding complex goals and constraints, citing website development and cloud resource planning as examples. These statements are Alibaba's capabilities, not results of independent comparisons.

For tasks such as structuring rules, plan verification, simulations, and creating 3D presentations, Alibaba advises using GLM-5.3 and DeepSeek-V4-Pro. For faster daily work, including text composition, information extraction, and content organization, it is recommended to use more budget-friendly options like Qwen3.8-Flash, GLM-5.3-Flash, and DeepSeek-V4.1-Flash. The announcement does not include a general benchmark determining which option is best suited for each task.

Enterprise account administrators can configure model access by department, user group, or individual member, according to Alibaba. This allows organizations to offer different options for different roles instead of showing the same menu to everyone.

This update followed the recent release of QwenWork 2.0 with team workspaces. Its immediate change concerns the level of model selection and permissions: the office agent can now utilize a wider range of providers, while the user or organization decides which models are available for specific work.

Alibaba integrates Qoder mechanism into Qwen Code 1.0 preview and opens agent core
Read more
pandaily.com

Alibaba integrates Qoder mechanism into Qwen Code 1.0 preview and opens agent core

Alibaba announced the release of Qwen Code 1.0 on November 11th, integrating the 'harness' mechanism developed in the Qoder coding product into the open-source Qwen Code project. This mechanism represents a layer that connects the model, its tools, and the execution process, and Alibaba states that developers can now incorporate it into their products.

The Qwen Code 1.0 preview is available in a separate branch of the QwenLM/qwen-code repository, while the existing stable branch remains unchanged. On GitHub, next to the current version 0.27, the preview release v1.0.0 from November 11th is displayed, and the project is licensed under Apache-2.0.

Architecture Details of Version 1.0

According to the description provided by Alibaba to IT Home, version 1.0 is based on an event-driven bus in process and memory, extending a unified event model for distributed operation. Data reception, planning, and execution processes are structured in layers, and the separation between 'brain and hand' ensures independent management of reasoning orchestration and tool execution. The goal is to transition from a single process serving multiple sessions to multiple nodes supporting many more sessions while maintaining existing entry points and return paths.

Practical Advantages of the New Version

Alibaba listed four practical benefits. Modules for reasoning, memory, and tools can be connected around events with minimal interference to the main execution flow. Access to the model is through a single entry point that centralizes authorization, configuration, and call logging. Sandbox resources are allocated and reused within the same permission and configuration boundaries, which, according to Alibaba, should reduce operating costs by decreasing reinitialization and downtime. Furthermore, the Model Context Protocol (MCP) for connecting tools is managed in one place for setup, authorization, and connections.

Qwen Code itself is an open-source coding agent that functions in the terminal. Its repository also demonstrates integrations for desktop, browser, editor, and chat, including applications such as VS Code, Zed, and JetBrains.

Alibaba has not yet published benchmarking results or a stable version 1.0 release date, and the announcement does not specify the degree of alignment between the open core and what operates inside the Qoder product. The 'preview' tag also indicates that interfaces may continue to change. The performance and cost-saving advantages described above are proprietary claims by Alibaba; the material reviewed did not contain independent testing.

Alibaba's T-Head chip division files for Hong Kong IPO
Read more
technode.com

Alibaba's T-Head chip division files for Hong Kong IPO

Alibaba announced on Tuesday that its division specializing in chip development, named T-Head, has confidentially filed documents for listing in Hong Kong on September 11.

The Hong Kong Stock Exchange has permitted Alibaba to proceed with the planned asset spin-off, but final approval for the listing has not yet been granted.

T-Head develops chips intended for data centers and is currently wholly owned by Alibaba. The group plans to retain control over the division after its public offering. Currently, the structure and size of the proposed share placement are being worked out.

Popular