Xing4.0-29B-A4B: promising 24 GB agentic MoE, not yet a boring install

China Telecom AI released Xing4.0-29B-A4B on September 16, 2026 under Apache 2.0. It is a 29B-total, 4B-active MoE with 256K context, trained on Ascend NPUs with MindSpore and aimed at coding and computer-use agents. The architecture is attractive for a 24 GB machine; the release is not mature enough to displace Qwen yet.

Verdict: an Apache-2.0 29B/4B-active agentic MoE with strong vendor scores — and a runtime stack still waiting on upstream PRs

The take

From the model card: 40 layers, 64 routed experts plus one shared expert, top-4 routing, MLA, mHC and MTP, with context extendable from 256K to 512K. The safetensor metadata totals about 31.2B parameters including embeddings. The vendor reports SWE-bench Verified 75.0, Terminal-Bench 2.1 57.5 and AIME 2026 90.0; these are vendor-run scores and have not been independently reproduced by us.

The deployment footnote is the reason for the verdict. The card names Transformers, vLLM, SGLang and KTransformers, but at release the required support lived in pending framework pull requests rather than ordinary stable installs. An official GGUF exists, but adoption is still small beside the base checkpoint.

Our call: watch. The licence, active-parameter count and long context are all good. Until the framework work lands upstream and community evaluations confirm the headline scores, Qwen3.8-27B and the established 30B-A3B family remain the safer planner recommendations.

Where this fits

Models: Qwen3.8-27B · Qwen 3.6-35B-A3B · Qwen3-Coder-30B-A3B

Hardware: NVIDIA RTX 5090 · NVIDIA RTX 4090 · M5 Max MacBook Pro 64 GB

Sources

Next step

Try this in the planner→