Xing4.0-29B-A4B is a text MoE model from China Telecom Artificial Intelligence Technology, developed as the next generation of TeleChat. The model has 29 billion parameters, of which approximately 4 billion are active per token. The developer designed it for complex engineering tasks and sequential agentic workflows.
Xing4.0 was trained entirely on Ascend NPU accelerators using MindSpore. Its architecture combines mHC, MLA and MTP; the model contains 40 layers, 64 routed experts, four selected experts per token and one shared expert. These details describe the underlying model’s architecture and research foundation, rather than the hardware processing your request.
The developer’s material emphasizes multistep planning, code work and sequential problem-solving. In a text conversation, use these capabilities to prepare a program-change plan, explain dependencies or investigate an error using supplied code. Start with the goal, input data and constraints, then ask for a sequence of actions and the expected outcome of each step.
For a programming task, provide the relevant implementation, data format and an example of expected behavior. A useful result includes an explanation, a localized change and example checks. After running them in your environment, return the observations to the conversation and refine the solution. This keeps the response connected to a specific task without attributing independent project access to the model.
In Ailoot API, Xing4.0 accepts text and returns text. Its context window is 262,144 tokens and its maximum output is 235,929 tokens. Instructions, source material, history and the response share the overall context. You can include several related passages: requirements, documentation, code and verification results. Give them clear names and specify the desired result—a comparison, an algorithm description, revised code or a solution plan.
The developer states that the underlying model’s context can be extended to 512K, but this connection provides 262,144 tokens. Use the latter limit when preparing a large request and leave room for the response.
Messages, Responses and Chat Completions are available. These formats help integrate the model into a familiar client and organize several successive messages. This connection returns generated text. Executing proposed code, reading files and working with external tools require the corresponding application logic. The underlying model’s architecture does not itself provide those actions within an ordinary text request.
", "aboutText": "### Xing4.0 from China Telecom AI\n\nXing4.0-29B-A4B is a text MoE model from China Telecom Artificial Intelligence Technology, developed as the next generation of TeleChat. The model has 29 billion parameters, of which approximately 4 billion are active per token. The developer designed it for complex engineering tasks and sequential agentic workflows.\n\nXing4.0 was trained entirely on Ascend NPU accelerators using MindSpore. Its architecture combines mHC, MLA and MTP; the model contains 40 layers, 64 routed experts, four selected experts per token and one shared expert. These details describe the underlying model’s architecture and research foundation, rather than the hardware processing your request.\n\n### Text-based engineering tasks\n\nThe developer’s material emphasizes multistep planning, code work and sequential problem-solving. In a text conversation, use these capabilities to prepare a program-change plan, explain dependencies or investigate an error using supplied code. Start with the goal, input data and constraints, then ask for a sequence of actions and the expected outcome of each step.\n\nFor a programming task, provide the relevant implementation, data format and an example of expected behavior. A useful result includes an explanation, a localized change and example checks. After running them in your environment, return the observations to the conversation and refine the solution. This keeps the response connected to a specific task without attributing independent project access to the model.\n\n### Context and request material\n\nIn Ailoot API, Xing4.0 accepts text and returns text. Its context window is 262,144 tokens and its maximum output is 235,929 tokens. Instructions, source material, history and the response share the overall context. You can include several related passages: requirements, documentation, code and verification results. Give them clear names and specify the desired result—a comparison, an algorithm description, revised code or a solution plan.\n\nThe developer states that the underlying model’s context can be extended to 512K, but this connection provides 262,144 tokens. Use the latter limit when preparing a large request and leave room for the response.\n\n### Connecting to an application\n\nMessages, Responses and Chat Completions are available. These formats help integrate the model into a familiar client and organize several successive messages. This connection returns generated text. Executing proposed code, reading files and working with external tools require the corresponding application logic. The underlying model’s architecture does not itself provide those actions within an ordinary text request.", "locale": "en", "methods": [ { "html": "POST /v1/messagestexttextPOST /v1/responsestexttextPOST /v1/chat/completionstexttext