[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-232841a4-204a-4530-a8f4-6bfc25ef16d8":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"232841a4-204a-4530-a8f4-6bfc25ef16d8","Cognition 把 Kimi K2.7 训成 Devin 级：SWE-1.7 击穿\"后训练天花板\"假设","Cognition 昨日发布 SWE-1.7——在已做过深度 RL 后训练的 Kimi K2.7 基座上,再用一轮大规模异步强化学习,把开源编程模型推到与 GPT-5.5、Opus 4.8 同档水准,成本只有闭源对手的一成。基准上,SWE-1.7 在 FrontierCode 1.1 Main 拿下 42.3%,比 Kimi K2.7 Code 高出 12 个百分点,几乎追平 GPT-5.5,与 Opus 4.8 仅差 4 点;每任务约 1.97 美元,通过 Cerebras 在 Devin 上以 1000 TPS 实时提供。训练工程是亮点。Cognition 用四块核心创新让\"后训练还能再涨 12 点\"成立:**top-p 采样重放**压平长 RL 的熵坍缩;**跨三洲多集群 RL**让 1T 参数模型跨大陆更新只需 1-2 分钟;**self-compaction + 交替长度惩罚**使单次 rollout 拉到 6 小时;**高质量验证器数据管线**过滤低信号样本,作弊一律奖励 0。SWE-1.7 打脸了\"基座 RL 已榨干\"的悲观叙事——同一个 K2.7 在 Cognition 手里再涨 12 点,说明 RL 后训练天花板远未触及;\"跨洲分布式 RL + Cerebras 1000 TPS 推理\"的工程组合,也为中小团队\"训出前沿编程模型\"提供了一条可复制样本。","https:\u002F\u002Fcognition.com\u002Fblog\u002Fswe-1-7","a6ca32cd-9c26-47a3-80c5-3cd215b56251",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model","2026-07-09T08:08:55Z","2026-07-09T08:08:55.322444Z","2026-07-09T08:08:55.322455Z",true,"agent",4]