Xing4.0-29B-A4B: an Apache-2.0 agent model that runs on one GPU
The 29B/4B-active MoE scores 75 on SWE-bench Verified, plugs into Claude Code and OpenCode, and was trained end-to-end on Huawei Ascend — no NVIDIA.
Copy markdown
SWE-bench Verified 75, on 4B active params
Xing4.0-29B-A4B is a 29B mixture-of-experts that fires only 4B parameters per token, yet posts 75.0 on SWE-bench Verified and 76.6 on Claw-Eval — agent-grade coding scores from a model small enough to fine-tune yourself. Apache-2.0 licensed, 256K native context (extendable to 512K), full weights on Hugging Face.
It fits on a single 16GB card
A low-bit quantized build lands around 15GB, so it runs on one RTX 4060 Ti-class GPU with no API in the loop. The team ships format adapters for Claude Code, OpenCode, OpenClaw and Hermes — point your existing agent harness at local weights and you have a self-hosted coding agent this weekend.
Trained end-to-end on Huawei Ascend
Xing4.0 was trained entirely on Ascend NPUs with MindSpore — no NVIDIA anywhere in the loop — and still lands agent scores near frontier open models. It's the clearest sign yet that competitive open weights can ship off a non-CUDA stack, which matters for anyone tracking GPU supply.
Elsewhere: Codex CLI 0.156.1 nudges you to Luna
OpenAI shipped Codex CLI 0.156.1: GPT-6 Sol and Luna are now in the model picker, and the rate-limit switch prompt recommends Luna when you hit a cap — a small default that steers heavy users onto the cheaper tier.