一张3090就能跑！全栈国产模型，把AI办公搬到企业本地
Ranking
Overall
61
Content
65
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - China Telecom has open-sourced Xing4.0-29B-A4B, a lightweight MoE model designed for private enterprise coding and agent workflows. Its 4-bit version can run on a single RTX 3090, making local AI deployment practical where data cannot leave company infrastructure.
- The model has 29B total parameters but activates about 4B per inference; 4-bit quantization reduces memory use to roughly 15GB.
- It supports 256K context, extendable to 512K, and uses MLA, multi-token prediction, and mHC connections to improve long-context efficiency, generation speed, and training stability.
- Training used Huawei Ascend 910C hardware and MindSpore/MindFormers, with reported system optimizations raising training throughput by about 96% over out-of-box performance.
- It targets repository-level coding and long-running agent tasks, with integrations for agent frameworks and inference stacks including OpenCode, Claude Code, SGLang, and vLLM.
Sources (1)
一张3090就能跑！全栈国产模型，把AI办公搬到企业本地
Public signals
N/A
TL;DR - China Telecom has open-sourced Xing4.0-29B-A4B, a lightweight MoE model designed for private enterprise coding and agent workflows. Its 4-bit version can run on a single RTX 3090, making local AI deployment practical where data cannot leave company infrastructure.
- The model has 29B total parameters but activates about 4B per inference; 4-bit quantization reduces memory use to roughly 15GB.
- It supports 256K context, extendable to 512K, and uses MLA, multi-token prediction, and mHC connections to improve long-context efficiency, generation speed, and training stability.
- Training used Huawei Ascend 910C hardware and MindSpore/MindFormers, with reported system optimizations raising training throughput by about 96% over out-of-box performance.
- It targets repository-level coding and long-running agent tasks, with integrations for agent frameworks and inference stacks including OpenCode, Claude Code, SGLang, and vLLM.