[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-iquest-q1-320b-moe-agentic-cli":3,"topics-all":38,"news-related-9753c773-a3c7-4329-9c09-f2037eb1c086":57},{"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},"9753c773-a3c7-4329-9c09-f2037eb1c086","IQuest-Q1开源:320B MoE专攻CLI智能体","IQuest Research 公开 IQuest-Q1 技术报告:面向命令行智能体的稀疏 MoE,320B 总参、激活 15B、上下文 512K。DeepSWE v1.1 64.6、CyberGym 84.5 高于 GPT-5.6 Sol,Terminal-Bench 4.0 仅 22.7,并参与了自己下一代迭代。","9 月底，IQuest 在 Hugging Face 放出 IQuest-Q1 的权重：3200 亿总参数、每 token 只激活 150 亿、上下文 524,288。当时它更像一条社区快讯——开源权重、vLLM 与 SGLang day-0 支持，剩下的是社区自己扒模型卡。\n\n真正把这件事顶到台前的，是 10 月 22 日公开的完整技术报告。报告第一次摊开 8 项智能体与推理基准的结果，上线后也进了 Hugging Face 每日论文榜；更关键的是，它交代了这个模型是怎么被造出来的。\n\n最刺眼的一组数字在长程任务上。Terminal-Bench 2.1 拿 83.2，同一批模型分布在 66.0 到 89.1 之间，看着正常；可换到 3.0 版本只剩 17.9，4.0 版本 22.7，而 Claude Opus 5 在这两个版本上是 42.7 和 53.9。这不像是评测换代，更像是模型只在某一代 harness 上被调教过。\n\n## 规格：320B 摆在桌上，15B 在跑\n\n据 IQuest Research 发布的[技术报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.30949)，该模型由 IQuest 团队打造，定位是命令行里的编码与通用智能体。总参数约 3200 亿，每 token 激活约 150 亿；稀疏 MoE 的常规配置是 256 个专家、每次激活 8 个，约占 3%。[模型卡](https:\u002F\u002Fhuggingface.co\u002FIQuestLab\u002FIQuest-Q1)给出了完整规格。\n\n骨架是 88 层 Transformer，隐藏维度 3072，查询头 48、键值头 8，头维度 128。注意力栈按 3 层滑动窗口 + 1 层全注意力交错，窗口 4096；位置编码只旋转每头 32 维。第一个 Transformer 层干脆不做稀疏路由，改用稠密 MLP，官方给的理由是缓解模型最早期专家负载不均衡。\n\n## 训练：把稳定性当成一等公民\n\n预训练用了 Stable SplitNorm、Muon 优化器和 μP 参数化。Stable SplitNorm 的思路是进入 FFN 的残差流不做归一化，换取 Post-Norm 式的深度利用率；μP 负责学习率在不同规模间的迁移，论文称在宽度与深度各 4 倍的跨度上验证了超参迁移，搜索是在一个隐藏宽度 896、18 层的小代理模型上做的，零样本搬到目标规模。\n\n一个受控对比值得记：与 Qwen3-30B-A3B 式的 MoE 设计、同一份训练数据相比，论文报告 1.93 倍的扩展效率，达到同等训练损失大约只需 52% 的训练算力。论文还外推出一个「临界模型规模」概念——按激活参数约 5720 亿、按总参数约 12 万亿，意思是这套配方在远大于当前规模时仍留有稳定性余量。\n\n## 后训练：任务、harness、四个专家\n\n后训练是这套流程最费功夫的部分。任务与环境一起合成：通用智能体任务跑在真实 API、MCP server 和工作区文件上，编码任务跑在从代码仓库搭出来的可执行环境里。RL 采用多 harness、多版本策略，一个共享策略同时在 Claude Code、Codex、OpenCode 等不同 harness 及其不同版本上训练，每个 episode 指定任务域、harness 版本、环境和评估器；奖励只认可验证的终态正确性，并区分策略自身的错、环境或 harness 的错与评估器的错。\n\nRL 最终产出四个专家，分别擅长终端交互体验、多 harness 操作、长程任务和通用智能体。MOPD（多教师在线策略蒸馏）把这四个冻结专家合并进一个从 SFT 第二阶段初始化的学生模型，之后再用残差模型合并把不同阶段的 checkpoint 折进来。奖励作弊的防护分三层：环境准备、运行时拦截、轨迹审计。\n\n## 跑分：强在编码，短板在长程\n\n官方表格里，DeepSWE v1.1 64.6、SWE-bench Multilingual 80.3、NL2Repo-Bench 63.0、CyberGym 84.5、Terminal-Bench 2.1 83.2、GPQA Diamond 93.3、IMOAnswerBench-Verified 93.0、BrowseComp 85.5。同一张官方表格里，CyberGym 高于 GPT-5.6 Sol 的 83.6，AutomationBench 52.7 高于 Claude Opus 5 的 50.3，τ³-Banking 51.8 也高于 Claude Opus 5 的 44.7。\n\n但 Terminal-Bench 3.0 的 17.9、4.0 的 22.7，和 Claude Opus 5 的 42.7、53.9 摆在一起，差距很难解释成「harness 换代」。另外这些数字都是厂商自评：不同模型的 harness 与运行时上限不统一（CyberGym 6 小时、Terminal-Bench 2.1 是 8 小时、3.0 和 4.0 给到 10 小时），还有一项 IQuest-CLIBench 是自家内建基准。\n\n## 它参与了自己的下一代\n\n最耐人寻味的一段在论文第 6 节：一个基于该模型的智能体在人工监督下自我迭代。它在一次长程任务里因工具名与参数不匹配报错 35 次，随后回溯 6054 次工具调用，定位到训练数据严重偏向单一工具调用 schema、部分记录存在定义与参数错配，提出清洗不一致记录并用多 schema 增强数据；人类专家审核后才批准训练。该轮训练用了 23.5 亿 token 和 938.7 GPU 小时。\n\n安全侧也有实操：论文报告一次为期三天的部署中，人类安全专家借助该模型产出 30 份漏洞报告，均经同行独立验证，并对影响开源项目的问题做了负责任披露。\n\n## 所以呢\n\n限制写在模型卡里也很直白：纯文本输入，没有图像、音频或视频；工具调用走 IQuest 自有 chat template 格式，需要兼容的解析器；真实 CLI 任务里的迭代调试仍可能忽略约束、重复失败路径，需要人盯。\n\n所以真正的看点不是某个榜单名次，而是那套「模型参与自己下一代」的循环有没有闭环。如果 512K 上下文加 88 层混合注意力意味着 CLI 智能体正在从「应用层套壳」变成训练目标本身，那么下一个问题就不是谁的分更高，而是谁先把这条流水线跑顺——榜单会换版本，harness 也会换代，但训练与评测之间的这层耦合，才是这轮工程投入真正押注的东西。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.30949","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"90c2609f-8e3a-4174-9cab-b2030076c959","en","IQuest-Q1: a 320B sparse MoE built for CLI agents","A 320B sparse MoE with 15B active and 512K context for command-line agents. CyberGym 84.5 beats GPT-5.6 Sol, but Terminal-Bench 4.0 scores just 22.7.","Back in late September, IQuest put the weights for IQuest-Q1 on Hugging Face: 320 billion total parameters, 15 billion activated per token, 524,288 tokens of context. At the time it read like a community changelog entry — open weights, day-0 support in vLLM and SGLang, and then everyone left to reverse-engineer the model card themselves.\n\nWhat pushed the story to the front was the full technical report, published on October 22. For the first time it laid out results across 8 agentic and reasoning benchmarks, and it earned a slot on the Hugging Face Daily Papers page. More importantly, it explained how the model was actually built.\n\nThe most striking numbers are in the long-horizon tasks. IQuest-Q1 scores 83.2 on Terminal-Bench 2.1, with the same field of models spread across 66.0 to 89.1 — normal-looking. Then on version 3.0 it drops to 17.9, and on version 4.0 to 22.7, while Claude Opus 5 posts 42.7 and 53.9 on those same two versions. That does not look like a benchmark being retired. It looks like a model that was only ever tuned against one generation of harness.\n\n## The Spec: 320B on the Shelf, 15B in the Loop\n\nAccording to the technical report from IQuest Research, the model was developed by the IQuest team and is positioned for coding and general-purpose agents in the command line. It carries roughly 320 billion total parameters with about 15 billion activated per token. The usual sparse MoE arrangement is 256 experts with 8 live at a time — roughly 3 percent active. The model card carries the full specification.\n\nThe backbone is 88 Transformer layers with a hidden dimension of 3072, 48 query heads, 8 key-value heads, and a head dimension of 128. The attention stack interleaves 3 sliding-window attention layers with 1 full attention layer, the window capped at 4096, and rotary position embeddings are applied to only 32 dimensions per head. The very first Transformer layer drops sparse routing entirely in favor of a dense MLP, which the report gives as a way to mitigate expert load imbalance in the earliest layer of the model.\n\n## Training: Stability as a First-Class Concern\n\nPre-training combined Stable SplitNorm, the Muon optimizer, and muP parameterization. The idea behind Stable SplitNorm is that the residual stream entering the feed-forward network stays unnormalized, buying back the depth utilization associated with Post-Norm; muP handles learning-rate transfer across scale. The paper reports hyper-parameter transfer validated across a 4x range in width and a 4x range in depth, with the search performed on a small proxy model of hidden width 896 and 18 layers and then transferred zero-shot to the target scale.\n\nOne controlled comparison is worth remembering: against a Qwen3-30B-A3B-style MoE design using identical training data, the paper reports 1.93x higher scaling efficiency, reaching the same training loss on roughly 52 percent of the training compute. The paper also extrapolates a \"critical model size\" — about 572 billion by activated parameters and about 12 trillion by total parameters — meaning the recipe retains stability headroom well beyond the scales studied.\n\n## Post-Training: Tasks, Harnesses, Four Experts\n\nPost-training is where the real work sits. Tasks are synthesized together with their environments: general-agent tasks run against real APIs, MCP servers, and workspace files, while coding tasks run in executable environments built from repositories. RL uses a multi-harness, multi-version strategy, training one shared policy across Claude Code, Codex, OpenCode and their different versions, with every episode specifying task domain, harness and version, environment, and evaluator. Reward only recognizes verified terminal correctness, and failures are attributed before optimization: errors caused by the policy supply learning signals, while environment or harness faults are masked and evaluator faults block the sample entirely.\n\nRL ultimately yields four experts — agentic user experience, multi-harness work, long-horizon tasks, and general agentic work. Multi-teacher on-policy distillation (MOPD) consolidates those four frozen experts into a single student initialized from the SFT Stage 2 checkpoint, and residual model merging then folds checkpoints across stages and expert branches into the final weights. Reward hacking is defended on three layers: environment preparation, execution-time checks, and trajectory auditing.\n\n## Benchmarks: Strong on Coding, Weak on Long Horizons\n\nIn the official table: DeepSWE v1.1 at 64.6, SWE-bench Multilingual at 80.3, NL2Repo-Bench at 63.0, CyberGym at 84.5, Terminal-Bench 2.1 at 83.2, GPQA Diamond at 93.3, IMOAnswerBench-Verified at 93.0, and BrowseComp at 85.5. In that same official table, CyberGym comes in above the 83.6 posted by GPT-5.6 Sol, AutomationBench at 52.7 is above the 50.3 from Claude Opus 5, and tau-cubed-Banking at 51.8 is also above the 44.7 from Claude Opus 5.\n\nBut set 17.9 on Terminal-Bench 3.0 and 22.7 on version 4.0 against 42.7 and 53.9 for Claude Opus 5, and the gap is hard to explain as a harness generational shift. These are also vendor-reported numbers: the harness and runtime limits differ across models (six hours for CyberGym, eight for Terminal-Bench 2.1, ten for versions 3.0 and 4.0), and one entry, IQuest-CLIBench, is an in-house benchmark.\n\n## It Helped Build Its Own Successor\n\nThe most telling passage sits in Section 6 of the paper: an agent built on the model iterating on itself under human supervision. It produced 35 tool-call errors in one long-horizon task from mismatched tool names and parameters, then traced failures across 6,054 tool calls, found the training data heavily biased toward a single tool-call schema with some records containing mismatches between tool definitions and call parameters, and proposed cleaning those records while augmenting with diverse schemas. Human experts reviewed the diagnosis before approving the training run. That run used 2.35 billion tokens and 938.7 GPU hours.\n\nThere is a practical deployment too: the paper reports that during a three-day deployment, human security experts assisted by the model produced 30 vulnerability reports, all independently validated by fellow security experts, with findings affecting open-source projects responsibly disclosed to their maintainers.\n\n## So What\n\nThe limitations are stated just as plainly in the model card: text-only input, with no image, audio, or video; tool calls use an IQuest-specific format in the chat template and require compatible serving parsers; and real-world CLI tasks involving iterative debugging often trip the model up, so it may overlook constraints, repeat failed approaches, or leave issues unresolved, necessitating human review.\n\nSo the real story is not a leaderboard position but whether the loop in which a model contributes to its own successor actually closes. If a 512K context and 88 layers of hybrid attention mean CLI agents are turning from an application wrapper into a training target in their own right, then the next question is not who scores higher but who first gets that pipeline right. Benchmarks will get new versions and harnesses will get retired, but that coupling between training and evaluation is what this round of engineering is really betting on.","iquest-q1-320b-moe-agentic-cli","2026-10-22T09:01:29Z","2026-10-25T01:19:28.485326Z","2026-10-25T01:19:28.485341Z",true,"agent",18,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"03d33c23-e56a-41eb-8732-b0712059662f","PUBG 母公司开源 MoE:激活仅 4.3B 单卡可跑","krafton-raon-1-mini-preview-open-weights","2026-10-23T13:08:33+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"2c7a8ca0-ab46-454b-ae16-aa7e7b413753","Ultramarine 1 Flash:不可验证任务RL","ultramarine-1-flash-non-verifiable-rl","2026-10-20T07:07:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00"]