China Telecom open-sources Xing4.0-29B-A4B, a 29B MoE agent model that runs on one RTX 3090
量子位 · wechat · 2026-09-20
China Telecom open-sourced Xing4.0-29B-A4B, a fully domestic-stack lightweight coding/agent model. The MoE has 29B total parameters with 4B active per token; 4-bit quantization cuts memory to 15GB, letting it run locally on a single RTX 3090. It's already #4 on HuggingFace trending.
Capabilities: native 256K context (expandable to 512K), repo-level coding and long-horizon agent tasks, with adaptations for OpenCode, Claude Code, vLLM, SGLang and more. In production it scores 200-page bid documents locally in 20 minutes and handles complex private-deployed customer service tasks at >90% success, all without data leaving the enterprise.
Technical details: MLA for KV-cache compression, MTP multi-token prediction, DeepSeek-style mHC connections; tens of trillions of tokens including agent trajectories; trained entirely on Ascend 910C with MindSpore at 96% higher throughput than out-of-box. Weights on GitHub, HuggingFace, ModelScope and others.
More from coding & agent
- 70 hands-on cybersecurity projects with full source code — tom_doerr · 2026-09-21
- Dev compares coding models building a coop game: V4.1 outshines Astra's 'pathetic' default taste — teortaxesTex · 2026-09-21
- 'Just 3 lines of code' backfires: dev argues tools should expose complexity, not hide it — willcb · 2026-09-21
- TypeSafe's Jev returns typed decisions with probabilities, not text — here's where it fits in agent loops — prakersh · 2026-09-21
- Jev Engineering gives agents a decision brain, 193x faster and 444x cheaper in tests — agihouse_org · 2026-09-21
- Dev Argues PAW Shouldn't Hide Its Complexity, Points to DSPy as the Better Playbook — willcb · 2026-09-21