29B total / 4B active (MoE)
262,144 tokens native; extensible to 524,288 tokens
Apache-2.0
Official open-weight, text-only, agent-oriented language model and the latest generation of the Xing series, formerly known as TeleChat. Xing4.0-29B-A4B uses a custom 40-layer Mixture-of-Experts architecture combining manifold-constrained Hyper-Connections (mHC), Multi-head Latent Attention (MLA), and Multi-Token Prediction (MTP). Its shipped configuration specifies a 3,584-dimensional hidden state, two initial dense layers, 64 routed experts plus one shared expert, four routed experts selected per token, 1,024-dimensional expert feed-forward layers, a 9,216-dimensional dense feed-forward size, mHC multiplicity four, and one next-token prediction layer. The model has 29B total parameters while activating approximately 4B per token. It natively supports 262,144-token contexts and is advertised as extensible to 524,288 tokens using YaRN-based position scaling. The model supports multi-step planning, structured tool calling, long-context reasoning, and switchable thinking through the enable_thinking chat-template option. China Telecom reports that Xing4.0 was trained entirely on Ascend 910C NPUs using MindSpore and MindFormers, making it the first model at this scale trained entirely on the domestic Ascend software and hardware stack. Fine-grained MoE communication optimization, selective recomputation, automatic graph and operator fusion, and custom fused mHC operators reportedly improved training throughput by approximately 96% over the unoptimized implementation. Vendor-reported scores include 75.0 on SWE-bench Verified, 66.0 on SWE-bench Multilingual, 57.5 on Terminal-Bench 2.1, 76.55 on Claw-Eval, 64.63 on Tau3-Bench, 60.8 on DeepResearchBII, 90.0 on AIME 2026, 69.67 on IFBench, and 61.0 on AA-LCR. The evaluations use long-running agent harnesses and unusually large inference budgets, including a 24-hour Terminal-Bench timeout, 210K-token SWE-agent contexts, and up to 131,072 output tokens on AIME. The BF16 repository totals approximately 62.4 GB across 41 safetensors shards, while official FP8 and GGUF variants were released concurrently. Transformers loading currently requires trust_remote_code=True because the checkpoint ships custom Xing4.0 configuration and modeling code. Official deployment support includes vLLM, SGLang, and KTransformers; fine-tuning support includes LLaMA-Factory and MindFormers; and the release reports targeted format alignment for OpenCode, Claude Code, OpenClaw, and Hermes.