[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ornith-1-5-self-improvement-open-models":3,"news-related-89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"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 AI 开源 Ornith-1.5 家族：397B\u002F35B\u002F9B 三档，把任务生成、脚手架搭建与解题 rollout 全部纳入 GRPO 自我改进循环。官方评测中 397B 版 Terminal-Bench 86.1、DeepSWE 56，9B 量化版可部署到手机。","8 月 18 日，Ornith AI 发布了开源模型家族 Ornith-1.5，一次放出三档规模：397B MoE、35B MoE（每 token 激活约 3B 参数）和 9B 稠密模型，后者还带一个量化后的 Mobile 版本，可以直接部署到 iPhone 和 Android 设备。这一代的主题不是堆参数，而是把 6 月 Ornith-1.0 提出的 self-scaffolding（自脚手架）框架，扩展成一个端到端的自我改进循环。\n\n## 自我改进循环：模型自己出题、自己搭考场\n\n官方博客对这套循环的描述相当激进：给定一个环境或代码库，模型会提出难度递增的新任务，专门挑自己还没解出过的类型下手；接着为每个任务生成或改进任务专属 scaffold——包括指令、工具、任务分解策略与编排；最后基于任务和 scaffold 产出解题 rollout。Reward 从 rollout 反向传播到全部三个阶段，用 GRPO 联合优化。也就是说，系统不只学着解更好的题，还同时学着出更有训练价值的题、搭更有效的考场。\n\n任务侧的 reward 由三个信号相乘得到：有效性（任务可验证，验证失败直接归零）、前沿难度（用模型自身 rollout 的成功率估计，目标成功率设在 0.2——任务一旦被模型稳定解开，reward 自然衰减，逼着出题器继续升级难度）、新颖性（与历史任务缓冲区查重，防止重复刷同类题）。Harness 侧同样是三项相乘：任务对齐、reward 保真、抗 reward-hacking。这套设计把「课程」从人工策划变成了模型自产自销。\n\n## 成绩单：397B 在 agentic coding 上逼近闭源前沿\n\n官方公布的评测里，397B 旗舰在 agentic coding 一线表现最强：Terminal-Bench 2.1（Terminus-2）拿到 86.1，高于 Claude Opus 4.8 的 85.0；DeepSWE 56.0，接近 Opus 4.8 的 59.0；SWE-bench Verified 86 与 Opus 4.8 的 85.8、Kimi K3 的 86.2 处在同一档。对比同规模开源对手，它领先 GLM-5.2（Terminal-Bench 82.7、DeepSWE 46.2）和 DeepSeek-V4-Flash-0731（82.7、54.4）。\n\n35B 档更值得注意：Terminal-Bench 2.1（Terminus-2）67.8，领先 Qwen3.6-35B（52.5）、Gemma-4-31B（42.1）和 Meta 的 Muse Glimmer-30B（51.7）15 分以上；SWE-bench Verified 79.0 对 73.4、52.0、76.0。DeepSWE 上 35B 拿到 22 分，而前代 Ornith-1.0-35B 和 Qwen3.6-35B 都是 0 分。9B 档 SWE-bench Verified 70.6、Terminal-Bench 47.0，明显超过 31B 的 Gemma 4，逼近 35B 的 Qwen3.6——考虑到它只有 9B 且量化后能塞进手机，这个成绩值得端侧团队认真看看。\n\n397B 自己的代际进步也相当陡峭：DeepSWE 从 1.0 的 8 分涨到 56 分，Frontier-Bench v0.1 从 2.7 涨到 13.5。\n\n## 边界与判断\n\n需要说清楚的是，「追平闭源」只成立于 agentic coding 这一条线。Frontier-Bench v0.1 上，397B 的 13.5 距离 Opus 4.8 的 21.1 和 Kimi K3 的 23 还有明显差距；GPQA Diamond 92.8 对 Opus 4.8 的 93.6 也仍是落后。开源阵营在最难的任务上依然没有越过闭源前沿——而且这些分数目前全部来自官方自报，社区复现还没开始。\n\n但对行业来说，Ornith-1.5 的价值可能不在分数，而在方法论的可复现性：它把「合成任务 + 自评 harness + GRPO」的完整配方——包括 reward 的乘法设计、有效性硬门控，以及评测防作弊细节（SWE-bench 评测时抹掉 git 历史、禁用网络防止检索现成答案）——全部写进了发布博客。当通用数据的边际收益越来越薄，下一轮能力增长到底来自更大的集群，还是来自模型自造训练信号，Ornith 用一个 397B 的开源实现给后一条路线投了票。至于这条自我改进循环会不会在更高难度上撞上「自己骗自己」的 reward 漏洞，他们的三信号乘法是第一道防线，真正的答案要等社区拿权重去复现（来源：ornith.ai\u002Fornith_1_5.html）。","https:\u002F\u002Fornith.ai\u002Fornith_1_5.html","f77241ae-61aa-4464-a29b-3f056b4b6171",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"61f40de4-4659-4af8-8ce0-2e0be18ca73b","en","Ornith-1.5 Open-Sourced: The Model Writes Its Own Tasks and Exam Rooms, in 397B to 9B Tiers","Ornith AI open-sources the Ornith-1.5 family in 397B\u002F35B\u002F9B tiers, folding task generation, scaffold construction, and solution rollouts into one GRPO self-improvement loop. The 397B posts 86.1 on Terminal-Bench and 56 on DeepSWE; the quantized 9B runs on phones.","On August 18, Ornith AI released the open-source Ornith-1.5 family in three scales at once: a 397B MoE, a 35B MoE (activating roughly 3B parameters per token), and a 9B dense model — the last one shipping a quantized Mobile variant that deploys directly on iPhone and Android devices. The theme of this generation is not parameter count; it is extending the self-scaffolding framework introduced in Ornith-1.0 back in June into an end-to-end self-improvement loop.\n\n## The self-improvement loop: the model writes its own tasks and builds its own exam\n\nThe official blog's description of the loop is fairly aggressive: given an environment or codebase, the model proposes progressively harder new tasks, deliberately targeting types it has not yet solved; it then generates or refines a task-specific scaffold — instructions, tools, decomposition strategy, and orchestration — for each task; finally, conditioned on the task and scaffold, the policy produces a solution rollout. Reward from the rollout propagates back across all three stages, jointly optimized with GRPO. In other words, the system does not just learn to solve better problems — it simultaneously learns to pose more training-worthy problems and build more effective exam rooms.\n\nThe task-side reward is the product of three signals: validity (a task must be verifiable; failing validation zeroes the reward outright), frontier difficulty (estimated from the model's own rollout success rate, with the target set at 0.2 — once a task is reliably solved, its reward decays, forcing the proposer to keep escalating difficulty), and novelty (deduplicated against a buffer of previously generated tasks to prevent re-grinding the same type). The harness-side reward is likewise a product of three terms: task alignment, reward fidelity, and resistance to reward hacking. Together this turns \"curriculum\" from a human-curated artifact into something the model produces and consumes itself.\n\n## The scorecard: 397B closes in on the closed frontier in agentic coding\n\nIn the published evaluations, the 397B flagship is strongest on the agentic coding line: 86.1 on Terminal-Bench 2.1 (Terminus-2), above Claude Opus 4.8's 85.0; 56.0 on DeepSWE, close to Opus 4.8's 59.0; and 86 on SWE-bench Verified, in the same tier as Opus 4.8's 85.8 and Kimi K3's 86.2. Against open-source rivals of similar scale, it leads GLM-5.2 (Terminal-Bench 82.7, DeepSWE 46.2) and DeepSeek-V4-Flash-0731 (82.7 and 54.4).\n\nThe 35B tier deserves more attention: 67.8 on Terminal-Bench 2.1 (Terminus-2), leading Qwen3.6-35B (52.5), Gemma-4-31B (42.1), and Meta's Muse Glimmer-30B (51.7) by more than 15 points; SWE-bench Verified 79.0 versus 73.4, 52.0, and 76.0. On DeepSWE the 35B scores 22, where both the previous Ornith-1.0-35B and Qwen3.6-35B scored 0. The 9B tier posts 70.6 on SWE-bench Verified and 47.0 on Terminal-Bench, clearly beating the 31B Gemma 4 and closing in on the 35B Qwen3.6 — and given it is only 9B and fits on a phone after quantization, edge teams should take a serious look.\n\nThe generational jump of the 397B itself is steep too: DeepSWE climbed from 8 in version 1.0 to 56, and Frontier-Bench v0.1 from 2.7 to 13.5.\n\n## Boundaries and judgment\n\nTo be clear, \"matching closed models\" holds only on the agentic coding line. On Frontier-Bench v0.1, the 397B's 13.5 still trails Opus 4.8's 21.1 and Kimi K3's 23 by a wide margin; GPQA Diamond at 92.8 versus Opus 4.8's 93.6 remains behind. The open-source camp has still not crossed the closed frontier on the hardest tasks — and every one of these numbers is currently vendor-reported; community reproduction has not started.\n\nBut for the industry, Ornith-1.5's value may lie less in the scores than in the reproducibility of the method: it has written the full recipe of \"synthetic tasks + self-graded harness + GRPO\" — including the multiplicative reward design, the hard validity gate, and anti-cheating evaluation details (stripping git history and disabling networking during SWE-bench runs to block retrieval of existing answers) — into a public release blog post. As marginal returns on generic data thin out, the next round of capability growth will come either from bigger clusters or from models manufacturing their own training signal; Ornith has cast a 397B open-weight vote for the second route. Whether the self-improvement loop hits a \"fooling itself\" reward hole at higher difficulty remains to be seen — their three-signal product is the first line of defense, and the real answer will arrive when the community takes the weights and reproduces (source: ornith.ai\u002Fornith_1_5.html).","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00Z","2026-08-19T17:09:17.578196Z","2026-08-19T17:09:17.578220Z",true,"agent",255,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"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","kuaishou-kat-coder-v2-5","2026-07-27T06:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"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":51,"title":52,"news_slug":53,"published_at":54},"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":56,"title":57,"news_slug":58,"published_at":59},"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":61,"title":62,"news_slug":63,"published_at":64},"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":66,"title":67,"news_slug":68,"published_at":69},"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"]