Xing4.0-29B-A4B

// Overview

Xing4.0-29B-A4B is an open-weight agent model from China Telecom Artificial Intelligence Technology Co., Ltd., the latest generation of the lab's TeleChat family. It is a mixture of experts: 29B total parameters with only 4B activated per token, a native 256K-token context extensible to 512K, and an architecture the project describes as mHC + MLA + MTP — Multi-head Latent Attention over 64 routed experts, 4 active per token plus one shared expert, across 40 layers. The weights landed on Hugging Face on 16 September 2026; the project's own news log dates the open-source release 17 September, alongside FP8 and GGUF packings.

The headline of the release is where it was trained rather than what it scores. China Telecom says Xing4.0-29B-A4B was trained entirely on the Ascend NPU platform — Ascend 910C clusters running MindSpore and MindFormers — and calls it the first model of this scale trained end to end there. The write-up credits fine-grained MoE communication optimisation, selective recomputation, DVM automatic graph-operator fusion and Ascend C fused operators for mHC with lifting overall training throughput roughly 96% over out-of-the-box performance.

The model is aimed squarely at agent work: multi-step planning, tool calling and long reasoning chains, with format alignment done specifically for agent harnesses including OpenCode, Claude Code, OpenClaw and Hermes. On China Telecom's own evaluation table it reaches 75.00 on SWE-bench Verified, 66.00 on SWE-bench Multilingual and 57.50 on Terminal-Bench 2.1, and 76.55 on Claw-Eval, 64.63 on Tau3-Bench and 60.80 on DeepresearchBII. Against the two peers the lab chose to publish, it leads Gemma4-26B-A4B on almost every row and trades with Qwen3.6-35B-A3B — ahead on Terminal-Bench 2.1, Claw-Eval and DeepresearchBII, behind on SWE-bench and AIME2026.

Deployment is open-source-ecosystem first. Transformers works for local inference, and vLLM, SGLang, KTransformers and TensorRT-LLM support was submitted upstream as pull requests that the README notes were still pending at release — so until those merge, serving needs an install from the relevant PR branch. Fine-tuning is supported through LLaMA-Factory and MindFormers, and multi-chip deployment through BAAI's FlagOS. Everything is Apache 2.0.

// Benchmarks

This model's scores

1. SWE-bench Verified75%
2. SWE-bench Multilingual66%
3. Terminal-Bench 2.157.5%
4. Claw-Eval76.55%
5. Tau3-Bench64.63%
6. DeepresearchBII60.8%
7. AIME202690%
8. IFBench69.67%
9. AA.LCR61%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Strong coding-agent scores for its activated size — 75.00 SWE-bench Verified and 57.50 Terminal-Bench 2.1 with only 4B parameters active per token
• Native 256K context, extensible to 512K, aimed at long multi-step agent runs
• Apache 2.0 weights published in BF16, FP8 and GGUF on Hugging Face, ModelScope and Modelers
• Trained entirely on Ascend 910C clusters with MindSpore — an existence proof for the non-NVIDIA training stack at this scale
• Format alignment done for agent harnesses such as OpenCode, Claude Code, OpenClaw and Hermes

// Best for

• Running a coding or research agent locally where an open-weight MoE with a small active footprint fits the hardware budget
• Long-context agent workloads that need 256K or more tokens of task state
• Fine-tuning for vertical domains — intent classification, table understanding, contract auditing, knowledge-based QA — on proprietary data
• Evaluating what a fully Ascend-trained model delivers against NVIDIA-trained open-weight peers of similar size