Xing4.0: open Apache-2.0 agent MoE, 4B active, SWE-bench 75

China Telecom's 29B model activates just 4B a token, ships GGUF and multi-NPU builds, and was trained end-to-end on Huawei Ascend instead of Nvidia.

Nowline SEP 19 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A 29B agent model that costs like a 4B

    Xing4.0-29B-A4B is a sparse MoE — 64 experts, 4 routed plus 1 shared, ~4B active per token — under Apache 2.0. You get near-frontier coding-agent behavior (SWE-bench Verified 75.0) at small-model inference cost and latency, with weights you host yourself.

  • 256K context, tuned for tool calls and long plans

    It ships a 256K native context window, extensible to 512K, on an mHC + MLA + MTP architecture built for multi-step planning, tool calling, and long reasoning chains — the workload for coding and terminal agents, not chat.

  • Trained end-to-end on Huawei Ascend, not Nvidia

    This is the first model at this scale trained fully on Huawei's Ascend NPU with MindSpore instead of CUDA, with ~96% training-throughput gains from custom optimization. For builders it's a credible open model born outside the Nvidia stack — supply diversity, not just another checkpoint.

  • Runnable tonight, but check the math scores

    FlagRelease already put out GGUF, W4A8, and Ascend/Hygon/Arm builds, and it slots into vLLM, SGLang, KTransformers, and LLaMA-Factory. It tops Terminal-Bench 2.1 (57.5 vs Gemma4's 30) and Claw-Eval (76.55), but trails on pure math (AIME2026 90.0 vs Qwen3.6's 92.7) — reach for it on agents, not olympiad problems.