{"id":85807,"topic":"ai","source":"手机新浪网","title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","url_hash":"0a0d20471df1ea62dedde63a308a58573fda999a","author":"","summary":"<a href=\"https://news.google.com/rss/articles/CBMidEFVX3lxTE04NkJkY0twa3FIaF9ySVBxZWFWUW14cEY4cHlJeGRybEdGeHhWWGtkY0VJSHR5eG4wSjNxSktPMmk4VFY2M0duOG1VQU40ZkNZUWRDTHhLMTczVnR4OGJROHc5ZWNqT2d2aHp5VE1hUTFma3ds?oc=5\" target=\"_blank\">中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">手机新浪网</font>","content":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","image_url":"https://n.sinaimg.cn/spider20260919/533/w820h513/20260919/8f4b-a015e015f5f68398e0006390b7dc04db.jpg","lang":"zh","published_at":"2026-09-19T11:13:33+00:00","fetched_at":"2026-09-19T11:15:09+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","cluster_id":3515894,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 608 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":608,"summary_length":608,"usable_text_length":608,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":608,"summary_length":608}},"news_item":{"id":85807,"canonical_url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","source_url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","source_name":"手机新浪网","author":null,"published_at":"2026-09-19T11:13:33+00:00","locale":"zh","topic":"ai","tags":[],"rss_summary":"<a href=\"https://news.google.com/rss/articles/CBMidEFVX3lxTE04NkJkY0twa3FIaF9ySVBxZWFWUW14cEY4cHlJeGRybEdGeHhWWGtkY0VJSHR5eG4wSjNxSktPMmk4VFY2M0duOG1VQU40ZkNZUWRDTHhLMTczVnR4OGJROHc5ZWNqT2d2aHp5VE1hUTFma3ds?oc=5\" target=\"_blank\">中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">手机新浪网</font>","full_text":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","excerpt":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 608 characters.","diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 608 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":608,"summary_length":608,"usable_text_length":608,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":608,"summary_length":608}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","summary":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","source":"手机新浪网","date":"2026-09-19T11:13:33+00:00","content":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 608 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 608 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":608,"summary_length":608,"usable_text_length":608,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":608,"summary_length":608}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/85807","export_markdown":"/api/items/85807/export?format=markdown","export_json":"/api/items/85807/export?format=json","diagnose":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4"},"formats":{"full":{"id":85807,"title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","source":"手机新浪网","author":null,"published_at":"2026-09-19T11:13:33+00:00","locale":"zh","topic":"ai","tags":[],"excerpt":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","full_text":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","reading_time_min":1,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 608 characters.","diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 608 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":608,"summary_length":608,"usable_text_length":608,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":608,"summary_length":608}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 608 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":608,"summary_length":608,"usable_text_length":608,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":608,"summary_length":608}},"actions":{"read":"/item/85807","export_markdown":"/api/items/85807/export?format=markdown","export_json":"/api/items/85807/export?format=json","diagnose":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4"}},"digest":{"id":85807,"title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","source":"手机新浪网","topic":"ai","published_at":"2026-09-19T11:13:33+00:00","excerpt":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B IT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。 官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。 从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 608 characters.","reading_time_min":1,"cluster_id":3515894},"card":{"display_title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","subtitle":"手机新浪网 · 2026-09-19","summary":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B IT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。 官方表示，从训练芯片到推理部署，全栈国产 ——…","badges":["quality:high"],"links":{"read":"/item/85807","original":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","diagnose":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4"},"quality_warning":null},"export":{"title":"中国电信开源首个全栈国产轻量级智能体大模型Xing4.0-29B-A4B|Qwen|框架|适配|训练|显存 - 手机新浪网","url":"https://finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html?fromtech=1&vt=4","summary":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","source":"手机新浪网","date":"2026-09-19T11:13:33+00:00","content":"中国电信开源首个全栈国产轻量级智能体大模型 Xing4.0-29B-A4B\nIT之家 9 月 19 日消息，中国电信于 9 月 17 日发布新一代星辰大模型 Xing4.0-29B-A4B，该模型总参数量 29B，激活参数仅 4B，原生支持 256K 上下文，可扩展至 512K，是国内首个基于国产算力与国产框架完成训练、面向复杂工程任务深度优化的百亿参数大模型。\n官方表示，从训练芯片到推理部署，全栈国产 —— 昇腾训练、国产框架适配、自研架构、开源生态，一条路走通。\n从 SuperCLUE 评测结果来看，在智能体能力方面，Xing4.0-29B-A4B 得分 93.52，位列第 3 名，与两款头部 Qwen 模型的差距均不足 1 分。\n据介绍，Xing4.0-29B-A4B 通过 4-bit 量化后显存占用降低至 15GB（比 FP16 省约 75%），RTX 3090、RTX 4090 这类 24G 消费级显卡就能本地跑长上下文任务。\nXing4.0-29B-A4B 延续星辰大模型的开源路线，对主流开源生态做了全面兼容：支持 LLaMA-Factory、MindFormers 等微调架构；适配 SGLang、vLLM、KTransformers 等主流推理框架；Agent 框架对 OpenCode、Claude Code、OpenClaw、Hermes 做了定向适配调优，深度对齐各框架格式规范。","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//finance.sina.cn/tech/2026-09-19/detail-iniskimy7459024.d.html%3Ffromtech%3D1%26vt%3D4","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 608 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 608 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":608,"summary_length":608,"usable_text_length":608,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":608,"summary_length":608}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}