On September 22, 2026, China Telecom AI (TeleAI) officially released Xing4.0-29B-A4B via press release, positioning it as a next-generation lightweight agentic large model. The headline specs are straightforward: 29 billion total parameters with only 4 billion activated per token, a Mixture-of-Experts architecture, and a 256K-token context window. With low-bit quantization and memory optimization, the model needs just 15 GB of VRAM — a single consumer-grade GPU can run it locally.
Beyond the specs: an agentic model already in production
Unlike many releases that stop at benchmarks, TeleAI led with deployment. Xing4.0-29B-A4B has been integrated into China Telecom's group-level customer service platform, handling complex inquiries through multi-step reasoning and tool-calling workflows; it is also deployed in "mid-screen" interactive service scenarios, providing real-time assistance in home-service environments. The official line: the model has proven improvements in first-contact resolution and handling efficiency in these real businesses.
On benchmarks, the release cites 75.0 on SWE-bench Verified, claimed to rank among the top in its parameter class. Functionally it targets "execution" rather than "Q&A": autonomously planning multi-step task paths, calling external tools, processing long-context inputs, and delivering structured outputs across engineering scenarios like code development, data analysis, and document processing.
Open ecosystem: simultaneous release on GitHub and Hugging Face
The model is published on GitHub and Hugging Face (org: XingChen-AGI), compatible with mainstream training, inference, and agent-development frameworks, and validated across domestic and mainstream AI chip platforms. TeleAI says larger Xing4.0-series models are coming in the following months.
Our take: lightweight agents are becoming their own track
Placed on the recent industry map, the trend is clear: in the same week, Xiaomi released the MiMo-V2.6 dual full-modality versions, also aimed at the agent execution layer (Xiaomi releases MiMo-V2.6 full-modality duo: pushing the agent execution layer to the extreme (/zh/article/2026-xiao-mi-fa-bu-mimo-v-2-6-quan-mo-tai-shuang-ban-ben-ba-agent-zhi-xing-ceng-zuo-j)). NVIDIA's Nemotron 3.5 Lightning follows the same 30B-total / ~3B-activated logic. Vendors are converging on the same insight: an agent's high-frequency execution steps don't need to wake every parameter each time — "small activation, large total" is becoming the standard answer.
TeleAI's differentiation isn't in the parameters but in its identity: for a telecom operator building models, the biggest card is ready-made production scenarios. A customer-service platform handling massive real ticket volumes daily is validation no lab benchmark can provide. While others talk benchmarks, TeleAI talks "already running in our own business for a while" — more persuasive for enterprises considering private deployment.
For developers and SMEs, the most direct value is cost: one consumer GPU, data never leaving the intranet, and you get an agent base that can read code repos, call tools, and handle 256K-token contexts. How much of that 75.0 SWE-bench score materializes in real engineering work remains for community evaluation.
