新智元报道
还记得9月底，我们拉着IQuest-Q1做的那几场实测吗？
6个路口的早高峰，给它管；双11仓库里8台搬运机器人的调度，交给它；24层写字楼的电梯调度，也让它试了一遍。
我们还让它写前端、做游戏，看看一个320B的开源大模型，到底能把多少事情真正做起来。
那篇文章发出后，一个问题其实还没讲透：
能在复杂环境里连续干活的基础模型，究竟是怎么练出来的？
现在，这个问题终于有了更完整的答案。
就在今天，至知创新研究院（IQuest Research）公开了IQuest-Q1的完整技术报告《IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems》。
模型还是我们此前关注的那匹开源「黑马」：
320B总参数、15B激活参数，重点面向代码开发、命令行交互和复杂Agent任务。
不过，这次值得关注的，已经不只是模型会做什么了。
新公开的技术报告，给出了30项Benchmark成绩，还把不少训练细节摆到了台面上。
Benchmark成绩
对照模型包括Claude Opus 5、GPT-5.6 Sol等。而我们更感兴趣的，是这些成绩背后的技术投入：
29T Tokens、超过100万个可执行任务实例、1.93×的架构规模化效率，还有一套专门防止模型「作弊」的训练机制。
从9月底的实测，到如今的完整技术报告，IQuest-Q1终于把更多底牌亮了出来。
这次，我们就来拆拆它。
此次技术报告公布了30项评测结果，涵盖代码生成、软件工程、终端操作、工具使用、复杂工作流以及通用推理等多个方向。
其中，AutomationBench与τ³-Banking的成绩非常值得多看几眼。
AutomationBench关注跨应用自动化能力。模型需要在不同系统之间调用工具、完成操作，并在多步骤执行过程中保持任务状态和约束。
τ³-Banking则引入复杂服务业务场景，考察智能体理解业务规则、调用工具和完成多轮任务的能力。
这些任务有一个共同特点：完成某个操作只是其中一步，模型还得让整个流程顺利走下去。
IQuest-Q1在两项评测中分别取得52.7和51.8分，均位居报告列示的对照模型首位。
此外，在SkillsBench、MCP-Atlas和Toolathlon-Verified上，模型分别取得74.1、78.4和75.0分，覆盖技能使用、工具接口理解及外部工具调用等能力。
在Claw-Eval和IQuest-CLIBench上，模型分别取得77.1和53.7分。
其中，IQuest-CLIBench是至知团队设计的内部评测，通过用户交互模拟、执行轨迹和环境状态检查，进一步考察CLI Agent的任务完成与交互表现。
至于此前开源时重点展示的代码能力，IQuest-Q1在NL2Repo-Bench和CyberGym上分别取得63.0和84.5分。
BrowseComp 85.5、WideSearch 81.8、GPQA Diamond 93.3等成绩，也进一步展示了模型在信息搜索和知识推理方面的能力。
整体看下来，这次报告补充了更多复杂工作流和工具使用方面的评测证据。
当然，Benchmark还不能回答真实应用中的所有问题。面对更复杂、更开放的环境，模型仍然可能出错，也可能需要人工介入。
但至少，我们可以进一步追问：这些能力，究竟是怎样「炼」出来的？
如果说训练一个代码模型，需要让它阅读和理解大量代码，那么训练一个能够独立执行任务的Agent基础模型，就得让它真正进入环境里做事。
写完代码，要运行。
运行出错，要排查。
调用工具返回了意料之外的结果，还得继续调整方案。
整个过程是连续的，并且会不断产生新的状态。模型必须学会处理这些连续发生的情况，并且如果出现了错误，模型也需要从这些反馈里学习。
IQuest-Q1技术报告披露，团队从代码仓库和已有智能体轨迹出发，构建了超过100万个可执行、容器化的任务实例。
这些任务涉及软件工程、终端操作、代码理解与测试、程序生成、网络安全和专业工作等多个领域，并配套相应的执行环境与验证方式。
智能体训练任务与验证体系
模型在任务环境中执行操作，系统再筛选有效、完整的执行轨迹，将其用于中期训练。
但环境越来越真实，另一个问题随之出现。
基础模型也会寻找捷径。
例如，在代码修复任务中，智能体可能通过查找已有答案、读取隐藏测试，甚至修改评测脚本获得奖励。
从评分结果看，任务似乎完成了；实际上，模型可能并没有解决原本的问题。
在强化学习中，这类现象被称为Reward Hacking（奖励作弊）。
为此，至知团队建立了三层防护机制：
• 环境准备：清理可能泄露答案的代码历史与参考文件，隔离任务环境。
• 执行时检查：限制模型访问隐藏测试、已有解决方案等不应获取的资源。
• 轨迹审计：检查实际执行的命令、工具输出与提交结果，剔除确认存在作弊行为的轨迹。
这套机制旨在提高奖励反馈与真实任务完成情况的一致性。
训练环境之外，IQuest-Q1还需要适应不同Agent工具的工作方式。
在后训练阶段，团队围绕Claude Code、Codex等执行框架开展训练，结合监督微调、领域强化学习与多教师在线策略蒸馏（MOPD），增强模型对不同工具接口和交互流程的适应能力。
四类专业教师模型分别覆盖终端交互体验、多执行框架适应、长程代码与网络安全任务，以及通用智能体能力。
后训练流程
对于CLI 模型而言，能否稳定地理解工具、处理反馈并完成任务，正在成为基础能力的重要组成部分。
除此之外，技术报告还披露了IQuest-Q1在基础模型架构上的一组实验结果：
1.93×。
这是团队在规模化效率对照实验中得到的结果。
在使用相同训练数据的条件下，与Qwen3-30B-A3B配置的MoE风格基线相比，IQuest-Q1模型达到相同训练损失所需的计算量约为基线的52%。
规模化效率与稳定性实验
需要注意，这一结果来自受控架构实验，衡量的是达到相同训练损失所需的计算量，并不直接等同于完整320B模型的实际训练成本。
那么，它背后有哪些设计？
IQuest-Q1采用320B总参数、15B激活参数的稀疏MoE架构。
模型包含88层Transformer，每个MoE层设置256个路由专家，每个Token激活8个。
注意力机制结合分组查询注意力与滑动窗口注意力，中间层采用每3层滑动窗口注意力搭配1层全注意力的结构。
IQuest-Q1整体模型架构
在优化方法上，团队采用Stable SplitNorm、Muon优化器与μP参数化，并通过小规模代理模型研究学习率迁移，改善大规模训练的效率和稳定性。
技术报告还引入Critical Model Size（CMS）指标，通过分析最优学习率和安全学习率之间的关系，研究模型在更大规模下的训练稳定性边界。
而这些架构设计，最终要承载庞大的训练任务。
报告显示，IQuest-Q1的预训练与中期训练累计使用约29T Tokens。
预训练阶段，代码数据比例由40%逐步提升至70%；中期训练进一步引入推理问答与智能体执行轨迹，并分阶段将最大训练上下文长度扩展至32K、128K和512K Tokens。
长上下文之外，团队还采用了一个有意思的推理设计：Recursive MTP（递归式多Token预测）。
模型在预训练时引入两个MTP层，之后将其中一层改造为递归草稿生成模块。
通过复用同一个MTP模块，系统能够在固定参数预算下生成多个后续候选Token，为投机解码提供支持。
此外，递归模块采用更小窗口的滑动窗口注意力，以限制长上下文条件下每一步预测的计算开销。
递归式多Token预测示意
从架构训练到推理执行，围绕计算效率问题，IQuest-Q1团队确实在多个环节进行了精心设计。
看完这份技术报告，再回头看我们上次让IQuest-Q1调度红绿灯、仓库机器人和电梯的实测，就更能理解一件事：
让模型持续完成复杂任务，确实没那么简单。
百万级可执行任务、多框架强化学习、防奖励作弊、长上下文推理……这些看似分散的技术设计，其实都在解决任务实际执行中可能遇到的问题。
报告里还有个有意思的细节。
在一次工具调用能力优化实验中，基于IQuest-Q1的智能体追查了6054次工具调用记录，发现训练数据过度集中于单一工具调用格式，部分工具定义与参数也存在不匹配。
随后，它提出数据清洗与增强方案，并在人工审核后参与后续训练准备。
连训练中遇到的问题，也开始尝试交给模型自己来排查了。
从技术报告里可以看到，研发团队是从预训练做起，跑通了完整的研发路径，而不是基于其他模型微调而来的。
这对于一家成立时间不长的机构来说，可见其在此方向发力的决心。
无论是模型后续表现，还是至知研究院的后续发展，它的成长性是值得期待的。
当然，技术报告能解释模型是怎么训练的，却无法替代真实环境中的检验。
毕竟，Benchmark上的高分只是一场考试。工具突然报错、任务越跑越长、环境不断变化，才是模型真正要面对的考场。
上次，我们看到了IQuest-Q1能干什么。
这次，知道了它是怎么「炼」出来的。
至于下一次，不妨把难度再往上加一点。
参考资料：
技术报告：https://arxiv.org/pdf/2610.30949
Blog：https://iquestlab.github.io/
Hugging Face：https://huggingface.co/IQuestLab/IQuest-Q1
GitHub：https://github.com/IQuestLab/IQuest-Q1
编辑：元宇
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Remember the fact that we pulled IQuest-Q1 at the end of September?
The early peaks at six junctions were piped; the dispatch of eight moving robots in the double 11 warehouse was handed over to it; and the elevator movement of the 24-storey writing building was also tested.
We also let it write the front end, play the game, see how much of a 320B open source LLM can really do.
After the article, one of the questions remains:
How did you get the basic model of working in a complex environment?
Now the answer to this question is finally more complete.
Today, the IQuest Research published the full technical report of IQuest-Q1: Advancing Foundations for Innovation CLI Systems.
The model is the same open-source "black horse" we used to focus on:
320B total, 15B activated parameters focused on code development, command line interactive and complex Agent missions.
But this time it's not just what models do.
The newly released technical report gave 30 Benchmark scores and placed many training details on the table.
Benchmark scores
Comparative models include Claude Opus 5, GPT-5.6 Sol, etc. And we are more interested in the technical inputs behind these achievements:
29T Tokens, more than 1 million missionable examples, 1.93x architecture size efficiency, and a training mechanism specifically designed to prevent the model of cheating.
From the facts at the end of September to the complete technical report today, IQuest-Q1 finally showed more ground cards.
This time, we'll tear it down.
The technical report published 30 assessments covering various directions, including code generation, software engineering, terminal operations, tool use, complex workflow and general reasoning.
Among them, Automation Bench and 3-Banking's achievements are worth looking at.
Automation Bench focuses on cross-application automation. Models need to call tools between different systems, complete operations and maintain mission status and constraints during multi-step implementation.
3-Banking, on the other hand, introduced a complex service business landscape to examine the ability of AI agent to understand business rules, access tools and perform multiple assignments.
These tasks have a common feature: the completion of an operation is only one step in the way, and the model has to keep the process going smoothly.
IQuest-Q1 achieved 52.7 and 51.8 points in two assessments, respectively, evenly ranked first in the comparison model presented in the report.
In addition, on SkillsBench, MCP-Atlas and Toolathlon-Verified, the models achieved 74.1, 78.4 and 75.0 points, respectively, covering skills use, tool interface understanding and external tool call capabilities.
In Claw-Eval and IQuest-CLIBench, the models scored 77.1 and 53.7 points respectively.
Among them, IQuest-CLIBench is an internal assessment designed by a well-known team to further examine CLI Agent ' s mission completion and interactive performance through interactive user simulation, implementation of trajectories and environmental state checks.
As for the pre-open source focus on code capabilities, IQuest-Q1 obtained 63.0 and 84.5 points on NL2Repo-Bench and CyberGym, respectively.
The performances of Brownecomp 85.5, WideSearch 81.8, GPQA Diamond 93.3 have further demonstrated the capabilities of models in information search and knowledge reasoning.
Overall, this report complements the more complex evidence of workflow and the assessment of the use of tools.
Of course, Benchmark is not yet in a position to answer all the questions in the true application. In the face of a more complex and open environment, models are still subject to error and may require manual intervention.
But at the very least, we can ask further: how did these capabilities come from?
If training a code model requires it to read and understand a large number of codes, then training an Agent base model that is capable of performing its tasks independently requires it to actually enter the environment to do its work.
Finish the code, to run.
Error running, to check.
The call tool returned unexpected results and continued to adjust programmes.
The whole process is continuous and creates new conditions. Models must learn to deal with these continuums and, if mistakes occur, they need to learn from these feedbacks.
The IQuest-Q1 technical report revealed that the team built more than 1 million examples of actionable, containerized missions from the code warehouse and the AI agent trajectory.
These tasks cover the areas of software engineering, terminal operations, code understanding and testing, programme generation, network security and professional work, and are accompanied by the corresponding implementation environment and certification modalities.
AI agent Training Mission and Certification System
The model operates in the mission environment and the system rescreens the effective and complete implementation trajectory for use in medium-term training.
But the environment is becoming more and more real, with another problem.
Basic models also look for shortcuts.
For example, in a code repair mission, AI agent may be rewarded by searching for available answers, reading hidden tests and even modifying the assessment script.
From the rating results, the task appears to have been accomplished; in fact, the model may not have solved the original problem.
In the context of intensive learning, such phenomena are known as Reward Hacking (reward for cheating).
To this end, the team has established three layers of protection:
• Environmental preparedness: cleaning up code history and reference documents that may reveal the answer, isolating the mission environment.
• Inspection at the time of implementation: limit model access to undesired resources such as hidden tests, solutions already in place.
• Track audit: examine the actual execution of orders, tool output and submission results and remove tracks confirming the existence of fraudulent behaviour.
This set of mechanisms aims to improve the consistency of reward feedback with the completion of real task.
In addition to the training environment, IQuest-Q1 needs to adapt to different Agent tools.
In the post-training phase, teams are training around the implementation frameworks of Claude Code, Codex and others to enhance the model ' s adaptability to different tool interfaces and interactive processes, in conjunction with monitoring fine-tuning, field-based enhanced learning and multi-teacher online strategy distillation (MOPD).
The four types of professional teacher model cover end-of-service interactive experience, multi-implementation framework adaptation, long-range code and network security tasks, and generic AI agent capability.
Post-training process
For the CLI model, the ability to understand tools, process feedback and fulfil mandates in a stable manner is becoming an important component of the base capacity.
In addition to this, the technical report discloses a set of IQuest-Q1 experiments on the underlying model structure:
1.93×。
This is the result of the team's scalability comparison experiment.
Under the same training data, the IQuest-Q1 model would need to calculate approximately 52 per cent of the baseline for the same training losses compared to the MoE-style baseline configured by Qwen3-30B-A3B.
Experiment on scale efficiency and stability
It needs to be noted that this result is derived from the controlled architecture experiment, which measures the calculation required to achieve the same training losses and does not directly equate to the actual training costs of the complete 320B model.
So, what are the designs behind it?
IQuest-Q1 uses a thin MoE structure of 320B overall, 15B activated parameters.
The model consists of 88 floor Transformer, with 256 routers per MoE layer and 8 activated per Token.
The attention mechanism combines the attention of the cluster query with the attention of the slide window, with the middle layer using a structure that combines the attention of every three floors of the slide window with one layer of full attention.
IQuest-Q1 overall model structure
In terms of optimisation, the team uses the Stable Splitnorm, Muon Optimizer and MiP parameters, and studies the migration of learning rates through small proxy models to improve the efficiency and stability of large-scale training.
The technical report also introduced the Critic Model Size (CMS) indicator, which examines the training stability boundaries of models at a larger scale by analysing the relationship between optimal and safe learning rates.
These structures are designed to eventually carry a huge training mission.
The report indicates that IQuest-Q1 has cumulatively used approximately 29T Tokens for pre- and medium-term training.
At the pre-training stage, the percentage of code data was gradually increased from 40 per cent to 70 per cent; the medium-term training further introduced the trajectory of reasoning questions and answers and AI agent implementation, and the maximum training context length was extended to 32K, 128K and 512K Tokens in phases.
In addition to long context, the team used an interesting reasoning design: Recursive MTP.
The model introduces two MTP layers during pre-training, and then transforms one of the layers into a draft-generation module of the Recursive.
By reusing the same MTP module, the system was able to generate several successor candidates, Token, under a fixed parameter budget, to support the decode of speculation.
In addition, the regression module uses a slided-window focus of a smaller window in order to limit the computational costs projected at each step under long context conditions.
Recursive Multiple Token Forecasts
From architecture training to reasoning implementation, the IQuest-Q1 team did indeed elaborate at multiple stages around computing efficiency.
After this technical report, and looking back at the last time we had IQuest-Q1 to move the red and green lights, warehouse robots and elevators, one thing is more understandable:
It's not that simple to keep the model on a complex mission.
One million-degree missionable, multi-framework-enhanced learning, incentive-proofing, long context reasoning These seemingly fragmented technical designs are in fact addressing the problems that may be encountered in the practical implementation of mandates.
There is also an interesting detail in the report.
In a tool call capability optimization experiment, Z42Z based on IQuest-Q1 traced 6054 tool call records and found that training data were overly concentrated in a single tool call format and that some tool definitions and parameters did not match.
It then proposes a data cleansing and enhancement programme and participates in the preparation of follow-up training after manual clearance.
Even the problems encountered during the training are beginning to be examined by the model itself.
As can be seen from the technical report, the R & D team started with pre-training and ran the full R & D path, rather than being fine-tuned based on other models.
This is the determination of an institution that has not been in existence for a long time.
Whether the model is followed by follow-up or by the Institute, its growth is to be expected.
Of course, the technical report explains how the model was trained, but is not a substitute for testing in the real environment.
After all, the high score on Benchmark was just an exam. It is the test field where models are really confronted with sudden errors in tools, longer missions and changing environments.
Last time we saw what IQuest-Q1 could do.
This time, I know how it came out.
For the next time, a little more difficulty might be added.
References:
Technical report: https://arxiv.org/pdf/2610.30949
Blog：https://iquestlab.github.io/
Hugging Face：https://huggingface.co/IQuestLab/IQuest-Q1
GitHub：https://github.com/IQuestLab/IQuest-Q1
Editor: Won-woo
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.