Xing4.0: a 4B-active open MoE that hits 75% on SWE-bench Verified
China Telecom's agent model runs at small-model speed, was trained entirely on Huawei Ascend NPUs, and now ships for nine chip families under Apache 2.0.
Copy markdown
A 4B-active MoE that scores 75% on SWE-bench Verified
Xing4.0-29B-A4B fires just 4B of its 29B parameters per token, yet posts 75.0% on SWE-bench Verified, 57.5% on Terminal-Bench 2.1, and 90.0 on AIME2026. It ships Apache 2.0 with a 256K context window that extends to 512K — a real local coding-and-agent brain, not a demo.
Only 4B active means it decodes like a small model
Because just 4B parameters light up per token, a Q4 GGUF runs at small-model speed while carrying 29B of knowledge. KTransformers even splits experts across CPU and GPU, so you skip the datacenter card. Point Aider or your own agent loop at a local vLLM endpoint this weekend.
Trained without a single Nvidia GPU
China Telecom trained the whole model on Huawei Ascend 910C NPUs with MindSpore, claiming roughly 96% higher training throughput from custom Ascend C operators and MoE communication tuning. It's the clearest proof yet that a competitive agent model can ship entirely off the CUDA stack.
Now runs on nine chip families, not just Ascend
Fresh FlagOS ports (build dated 2026-09-18) package the model for NVIDIA, Ascend, Moore Threads, MetaX, Hygon, Iluvatar, Tsingmicro, T-Head and ARM — BF16 across the board, W4A8 on ARM. 'Develop once, run anywhere' means your deployment isn't chained to one vendor's silicon.
How to run it this weekend
Serve it with vLLM (MTP speculative decoding), SGLang (EAGLE), or KTransformers for tight VRAM; grab the official FP8 or GGUF builds and fine-tune via LLaMA-Factory or MindFormers. The 256K native context comfortably swallows a mid-size repo for whole-project agent runs.