[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kuaishou-kat-coder-v2-5":3,"news-related-1e7d0673-aecc-42b5-8560-92a2b4d4daf6":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"1e7d0673-aecc-42b5-8560-92a2b4d4daf6","快手 KAT-Coder-V2.5 把 Agentic Coding 训练改写成基础设施工程","快手 KwaiKAT 团队 7 月 26 日发布 KAT-Coder-V2.5,在统一 Claude Code harness 下 PinchBench 拿到 94.9 分,首次压过 Claude Opus 4.8 的 93.5;同时把 35B 总参 \u002F 3B 激活的 MoE 变体 KAT-Coder-V2.5-Dev 以 Apache-2.0 推到 Hugging Face。\n\n但比刷榜更值得拆的是训练栈底下的工程改造。KwaiKAT 最初把 RL 曲线的迟缓归罪到算法本身,审计后才发现 16% 的轨迹失败来自 sandbox,不是策略。沙箱镜像把磁盘用到 95%,让超时 invalid rollout 占 6%–7%;远端 sandbox 初始化的环境变量错位再误翻 6%–7% 的 reward;Gateway Server 套的 chat 端点在 ~200 turn 规模上制造 40% 的 token drift。三项基础设施改动叠加之后,sandbox 反馈错误率从 16% 压到 2% 以下,训练塌陷减少一个数量级。\n\n支撑这层的是 AutoBuilder:把可执行环境构建成功率从 16.5% 拉到 57.2%,堆出 10 万+ 个跨 12 种语言的可验证仓库任务,每个任务都是「golden patch + test patch + 自动生成的三段式描述」三元组,验证不读 exit code,而是解析结构化测试输出,要求 ≥90% 的预期测试在多次运行中复现。数据飞轮也不只过滤「最终 pass」的轨迹,而是给「接近 pass」的任务加 process-level hints,让 0 通过子集回到 ~20% 通过率,再以 hint-free 重放洗掉泄漏。\n\n开源变体 KAT-Coder-V2.5-Dev 的底子是 Qwen3.6-35B-A3B 加 127K SFT 与 RL 后训练。这意味着 agentic coding 能力可以以 35B 总参、3B 激活的 MoE 形态在单卡 4-bit 量化下跑通,Qwen 系开源生态又多了一块拼图。Agentic coding 的下一程竞争,不在模型多大,而在沙箱稳不稳、数据过不过滤、reward 会不会被 hacking。KAT-Coder-V2.5 的真正卖点不是 PinchBench 那一分,而是把 RL 训练从「算法题」改写成了「基础设施题」。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.05471","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"db74f7bf-a9df-420e-8d5a-8a9637fdde4f","en","KAT-Coder-V2.5: agentic coding as infrastructure","On July 26, Kuaishou's KwaiKAT team released KAT-Coder-V2.5, scoring 94.9 on PinchBench under a unified Claude Code harness — the first time it edged out Claude Opus 4.8 (93.5) — and simultaneously pushed a 35B-total \u002F 3B-active MoE variant, KAT-Coder-V2.5-Dev, onto Hugging Face under Apache-2.0. But more worth dissecting than the leaderboard win is the engineering overhaul underneath the training stack. KwaiKAT originally blamed the slow RL curve on the algorithm itself, but an audit found that 16% of trajectory failures came from the sandbox, not the policy. Sandbox images were filling disks to 95%, making timeout-driven invalid rollouts account for 6%–7%; remote-sandbox initialization env-var mismatches cost another 6%–7% of reward; and the Gateway Server's chat endpoint introduced 40% token drift at the ~200-turn scale. After these three infrastructure fixes stacked up, the sandbox feedback-error rate dropped from 16% to under 2%, and training collapses shrank by an order of magnitude. Underpinning this layer is AutoBuilder: it pushed the executable-environment build success rate from 16.5% to 57.2%, stacking 100,000+ verifiable repository tasks across 12 languages, each as a \"golden patch + test patch + auto-generated three-segment description\" triplet. Verification doesn't read exit codes — it parses structured test output and requires ≥90% of the expected tests to reproduce across multiple runs. The data flywheel doesn't just filter \"eventually pass\" trajectories — it adds process-level hints to \"near-pass\" tasks, pushing 0-pass subsets back to ~20% pass rates, then washes away leaks via hint-free replay. The base of the open-source KAT-Coder-V2.5-Dev is Qwen3.6-35B-A3B with 127K SFT and RL post-training. This means agentic coding capability can now run, in a 35B-total \u002F 3B-active MoE form, in 4-bit quantization on a single card — and the Qwen open-source ecosystem has another piece of the puzzle. The next lap of agentic coding won't be won by model size — it'll be won by sandbox stability, data hygiene, and reward hacking resistance. KAT-Coder-V2.5's real selling point isn't that one PinchBench point — it's rewriting RL training from \"an algorithm problem\" into \"an infrastructure problem\".","kuaishou-kat-coder-v2-5","2026-07-27T06:00:00Z","2026-07-27T00:06:24.213127Z","2026-08-19T02:08:40.142862Z",true,"agent",109,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f","Ornith-1.5 开源：模型自己出题、自己搭考场，397B 到 9B 三档齐发","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"6ceaf229-f1a2-4231-b531-797a99faa194","Ornith-1.0：模型自写 RL harness，SWE-Bench 比肩 Opus 4.7","ornith-1-0-397b-moe-swe-bench-opus-4-7","2026-06-26T18:01:01+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"d63524e2-85bc-484c-b4e7-7fac32c3ac08","GLM-5.2 即将全量上线 Coding Plan：智谱把\"编程开源\"卷成新一轮标配","glm-5-2-coding-plan-zhipu-open-source","2026-06-13T07:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"747713d0-690f-45c7-afb0-7d6e16cb2a33","Cohere North Mini Code 开源：30B MoE、3B 激活，单卡 H100 跑起 Agentic Coding","cohere-north-mini-code-30b-3b-h100","2026-06-11T12:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","MiniMax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00"]