OPEN LINK ↗
// 1h agoMODEL RELEASE
China Telecom AI launches Xing4.0-29B-A4B agent MoE
China Telecom AI has released Xing4.0-29B-A4B, an open-weight 29B Mixture-of-Experts model with 4B active parameters optimized for agentic workflows. Trained on Huawei Ascend NPUs, the model features a native 256K context window and out-of-the-box support for runtimes including vLLM and SGLang.
// ANALYSIS
The practical battleground in open-weight AI is shifting from parameter inflation to hyper-efficient 3B–5B active parameter MoEs that deliver reliable agentic reasoning on edge hardware.
- –**Local Agent Sweet Spot:** Activating only 4B parameters per token enables rapid inference and low memory overhead while preserving the multi-step planning and tool-calling capabilities needed for autonomous workflows.
- –**Hardware Independence:** End-to-end training on Ascend NPUs via MindSpore proves that viable frontier-class MoE architectures can be produced without reliance on the NVIDIA software and hardware stack.
- –**Expansive Context Window:** A native 256K context (scalable to 512K) makes the model directly useful for full-repo code ingestion and lengthy execution traces.
- –**Immediate Ecosystem Readiness:** Native integration with vLLM, SGLang, and popular agent frameworks significantly reduces developer friction for local and self-hosted deployments.
// TAGS
aillmmoeopen-sourcemachine-learningchina-telecomxingchenascendagents
DISCOVERED
1h ago
2026-09-18
PUBLISHED
1h ago
2026-09-18
RELEVANCE
8/ 10
AUTHOR
honozcom