China Telecom's Xing4.0-29B-A4B open-sourced: trained on Ascend NPUs, 75 on SWE-bench Verified
jacek2023 · reddit · 2026-11-07
China Telecom AI released Xing4.0-29B-A4B (GGUF quant by bartowski), a 29B-total/4B-active MoE with MLA attention and native 256K context (extensible to 512K) — reportedly the first model at this scale trained entirely on Ascend NPUs with MindSpore, with co-optimization boosting training throughput 96%. Agent-oriented architecture (mHC + MLA + MTP) targets multi-step planning and tool calling. Benchmarks: SWE-bench Verified 75.0, Terminal-Bench 2.1 57.5, Claw-Eval 76.55. Supports LLaMA-Factory/MindFormers fine-tuning and SGLang/vLLM/KTransformers deployment, with adaptations for agent frameworks like OpenCode and Claude Code.
More from Infra
- Dev builds bash script to email Sam Altman daily about his 'RAM idle 95%' claim — IainKay · 2026-11-07
- Quantized Qwen3 beats small full-precision model on a 16GB laptop across 200 tasks — Nir777 · 2026-11-07
- Epoch AI: Huawei's largest China AI data center holds just 7% of Colossus 2's compute — Jsevillamol · 2026-11-07
- Musk says SpaceX will launch StarMind orbital supercomputers from Q3 2027, co-designed with NVIDIA — XFreeze · 2026-11-07
- Netflix builds ontology-driven observability: telemetry, incidents and services as one knowledge graph — JeremyCMorgan · 2026-11-07
- Google buys Crusoe's Wyoming data center site as hyperscaler takeover of neoclouds begins — SumitGup · 2026-11-07