磐石·科学基础大模型2.0

• 1. overview (893 chars)

  The Panshi·ScienceOne Omni large model 2.0 (English name ScienceOne Omni) is a large scientific model jointly created by multiple research institutes organized by the Chinese Academy of Sciences. It was officially released at the World Artificial Intelligence Conference on July 17, 2026. It is not another general chat model, but an intelligent base that serves scientific research tasks. It focuses on "universal-specific integration" - putting general understanding abilities and professional abilities in various disciplines together in the same model for joint training. Officials said that in the evaluation of more than 60 professional scientific research tasks, Panshi 2.0 significantly surpassed general flagship models such as Gemini-3.1-pro and GPT-5.5 in most tasks, and reached the international best in tasks such as chemical property prediction, spectrum to molecular structure prediction, and protein site prediction. The model has been fully open source and can be accessed through the official website scienceone.ia.ac.cn.

• 2. Product background (1732 chars)

  Panshi is led by the Chinese Academy of Sciences and is jointly built by coordinating interdisciplinary resources. The Institute of Automation of the Chinese Academy of Sciences is the leading unit in technological research. The head of the Panshi team is Zeng Dajun, deputy director of the Institute of Automation. This is not a project of a single laboratory, but an institutionalized collective research: Institute of Mathematics and Systems Science, Institute of High Energy Physics, Institute of Mechanics, Institute of Geographic Sciences and Natural Resources, Tibetan Plateau Institute, Institute of Zoology, National Astronomical Observatory, Shanghai Ceramics Institute, China Science and Technology Units such as the Academy University and the Hong Kong Innovation Institute are responsible for the construction of vertical models and professional agents in the field. The industrialization platform Zhongke Wenge focuses on basic large-scale model training. Zhongke Zidong Taichu is responsible for tool bases and intelligent agent factories. Enterprises such as Huawei, Haiguang, and Silang Technology provide computing power support.   Judging from the timeline, Panshi’s iterations are very intensive. Version 1.0 was first launched at the World Artificial Intelligence Conference on July 26, 2025. It was fully open source at that time and was positioned as an "AI + scientific" operating system. V1.5 was released in December of the same year, which strengthened multi-modal reasoning and added capabilities such as innovation evaluation and intelligent agent factory. The "Panshi 100" model system was launched on April 28, 2026, building the three scientific modal bases of waves, spectra, and fields, and building 8 dedicated model clusters in scientific fields, covering more than 100 scientific research scenarios. 2.0 on July 17, 2026 is a comprehensive upgrade after intensive iterations this year. As of the release of 2.0, the first-generation product has served more than 30,000 scientific researchers.

• 3. Product functions and usage experience (2813 chars)

  The core of Panshi 2.0 is a progressive three-layer architecture, which is officially summarized as a complete scientific intelligence link of "unified encoding of scientific data - alignment of natural world knowledge - directional decoding of domain tasks". In layman's terms, it means to "read" various scientific modal data such as waves, spectra, fields, etc. in a unified manner first, then align it with the knowledge of the natural world, and finally provide answers to tasks in specific fields. One model can complete a variety of tasks such as interdisciplinary data understanding, knowledge reasoning, scientific prediction, and professional content generation. It attempts to solve two old problems at the same time: high subject barriers in traditional field models and insufficient professional scientific capabilities in general models.   Data is its backbone. The Science and Technology Corpus of the Chinese Academy of Sciences has collaborated with various research institutes to construct 8 million pieces of high-quality scientific reasoning data, covering more than 200 scientific research tasks. The Panshi team transforms scientists' way of thinking into reasoning principles, builds verifiable reasoning links, and allows the model to learn to reason based on knowledge and evidence chains. Officials summarize this set of effects as "high credibility, strong logic, and interpretable" - for scientific research scenarios, being able to trace evidence is more important than having beautiful answers.   Based on the model, Panshi has built an integrated intelligent scientific research platform that connects "data base - tool base - agent hub - heterogeneous computing power". The platform brings together more than 8,000 professional scientific research tools and skills libraries, and scientific researchers can customize special models and agents on demand. Several signature agents have been implemented on a large scale: the Literature Compass agent can generate professional-level reviews within 3 hours, with an accuracy of 90% in evidence attribution. It has generated nearly 40,000 reviews and accessed 170 million scientific documents, compressing literature research that used to take 3 to 5 days to 20 minutes; the innovation evaluation agent is responsible for identifying key issues and research directions; the agent factory supports automatic matching and scheduling of scientific research tools using natural language, and strings together complex scientific research processes for integrated execution.   In terms of computing power, Panshi 2.0 adopts a trinity heterogeneous system of "super computing + intelligent computing + fast computing". Super computers run general scientific calculations and high-throughput simulations, intelligent computing clusters guarantee training and inference, and fast calculations rely on dedicated scientific computing hardware such as Silang Technology’s “Sky Dome” to specifically complete high-precision long-term simulations such as molecular dynamics and quantum chemistry at extremely high speeds. The entire set of computing power has been adapted to domestic chips such as Ascend and Haiguang. Officials emphasize that this is a new scientific computing base that is “independent and controllable”.

• 4. Pricing and business model (733 chars)

  Panshi is taking the open source route, and the model has been fully open source. The official website scienceone.ia.ac.cn can be directly accessed and used. This is essentially different from the commercial closed source model. Its business logic is not to make money through API subscriptions, but as a national-level scientific research infrastructure, supported by projects such as the National Artificial Intelligence Application Base in the Scientific Field, the Chinese Academy of Sciences' Class A Pilot Project, and the Beijing Science and Technology Plan Innovation Project, with the goal of serving the scientific research community and building an "AI + science" ecosystem. The target users are very clear: frontline scientific researchers, scientific research institutes, universities and related enterprises, rather than ordinary consumers.

• 5. User reviews (1818 chars)

  Judging from the scale of implementation, Panshi has gained practical recognition from the scientific research community. The first generation has served more than 30,000 scientific researchers, and 2.0 has been implemented in more than 50 research institutes within the Chinese Academy of Sciences, more than 30 scientific research units in universities outside the academy, and central state-owned enterprises, and has gone global through the United Nations Educational, Scientific and Cultural Organization and the "Belt and Road" international scientific organization alliance. Wang Ting, an engineer at the Institute of Automation, mentioned that scientists from the second Qinghai-Tibet scientific expedition team once asked Panshi a professional question. Compared with other general models, the answer given by Panshi was more accurate and rigorous. After repeated verification, it was highly consistent with the actual scientific expedition situation, and was affirmed by scientific expedition experts.   The feedback on specific scenarios is also relatively solid. In the field of mechanics, the team developed fluid high-throughput computing software based on Panshi, pushing fluid simulation from "grid generation, iterative solution, and long-term verification" to "grid-free, second-level response", with results within 10 seconds and an error controlled within 5%. In the field of astronomy, the astronomical spectrum intelligent agent improves the identification accuracy of rare celestial objects by about 50% compared with the existing best methods. It has been deployed to the National Astronomical Observatory to serve the Guo Shoujing Telescope. In the field of chemistry, Panshi drives the machine scientist system of the University of Science and Technology of China to realize a closed loop of on-demand material creation. It has completed the synthesis and characterization of 462 sets of samples and tens of millions of virtual screenings. 8 of the top 20 candidate materials have better performance than the original optimal samples, and the system has been running stably for more than 1,500 hours.

• 6. Industry evaluation (1235 chars)

  The industry generally regards Panshi as an important breakthrough in China's "AI for Science" direction. It fills a real gap: no matter how strong the general-purpose large model is, its reliability and interpretability are often insufficient in professional scientific research tasks; while traditional domain-specific models are independent and difficult to collaborate across disciplines. Panshi's "universal-specialist" and unified scientific research platform are regarded as providing an interdisciplinary "operating system" for "artificial intelligence + science". The statements of many scientists at the press conference also echoed this trend - physical intelligence will transform AI from a scientific "reader" to a "creator" of knowledge, and the industry's focus will shift from "how stunning the demonstration is" to "how reliable the operation is."   It is worth noting that Panshi is not an isolated case, but part of China’s scientific and intelligent ecology. During the same period, there were also special computing models represented by the DPA large atomic model (DPA4 ranked first on the international list of materials discovery), as well as the "Bohr Scientific Research Space Station" platform that brought together more than 200 million documents and more than 50,000 scientific research tools. These resources are being coordinated and integrated to form a more complete scientific research intelligent ecosystem.

• 7. Controversies and Risks (1422 chars)

  What needs to be viewed calmly is the context of the official data. "More than 60 evaluations significantly surpassed Gemini-3.1-pro and GPT-5.5." This conclusion comes from the developer's self-evaluation and press conference. The selection and task scope of the evaluation set are biased towards scientific and professional tasks, and cannot be simply equated to the overall leadership of comprehensive intelligence - Panshi is not good at this in general conversations and daily tasks, and it was not originally designed for these scenarios. Third-party independent reproduction and horizontal evaluation are currently relatively limited.   Secondly, the value of Panshi is highly dependent on ecology and supporting facilities. Many of its highlights (document compass, fluid software, machine scientist system) are essentially a combination of "model + platform + dedicated hardware + domain data". Without the institutional resources of the Chinese Academy of Sciences, domestic computing power and massive scientific corpus, it is difficult for ordinary teams to replicate the same effect. For general researchers outside the hospital, the depth of use may be different from that of units within the hospital.   Third, it is highly bound to domestic computing power and independent controllable routes, which is a strategic advantage, but it also means that its positioning is more biased towards national scientific research infrastructure rather than a universal tool readily available to global developers. Although open source has lowered the threshold, it still requires considerable scientific background and computing power to truly achieve its results.

• 8. Applicable groups and usage suggestions (1026 chars)

  Panshi 2.0 is most suitable for people with clear scientific research tasks: researchers in scientific research institutes and universities, teams that need to do literature reviews and interdisciplinary data analysis, and scientists who are engaged in calculations and simulations in the fields of materials/chemistry/astronomy/mechanics/life sciences. For them, the literature compass, tool scheduling, and domain-specific models are tools that can actually improve efficiency, especially in scenarios where literature research is reduced from several days to 20 minutes.   It is not suitable for ordinary users who are looking for a general chat assistant, daily office work or content creation - Panshi is not for this, and it is not cost-effective to use it for writing copywriting and chatting. Researchers who want to try it can start directly from the official website scienceone.ia.ac.cn, first use low-threshold agents such as the literature compass and tool dispatcher, and then access the corresponding special models according to subject needs. If your field happens to be within the scope of mathematics, physics, chemistry, and science covered by Panshi, the benefits will be more obvious.

• 9. Summarize (528 chars)

  The Panshi·Science Basics Model 2.0 is a significant collective research effort on China's "AI for Science" route. Its real value does not lie in outperforming a certain general model, but in using a "universal-specific" model plus an integrated scientific research platform to build a traceable, collaborative, autonomous and controllable intelligent base for scientific research. Its ceiling depends on how wide the ecology can be spread and how deep it can be used outside the hospital, but the direction is clear and the implementation is solid, which deserves continued attention from the scientific research circle.

No in-depth report yet

• 

  rONALD181

  —

  中科闻歌主攻训练，紫东太初做工具底座，几十家研究所高校企业一起上，这种跨体制的创新联合体模式在国内AI圈还挺少见的。

---

• 

  SatsChasera76

  —

  已经落地50多家院内所30多家院外单位，还通过教科文组织走出去了，生态铺得比我想的快。

---

• 

  MariaFrank

  —

  文献罗盘是真的省事，以前做个综述光前期查文献起码耗一周，还得自己一篇篇读一篇篇记，现在两三个小时就能出一版图文表都有的初稿，重点是它还会把每篇文献的核心创新点标出来、给出证据归因，改改就能直接用，对我们这种天天泡文献的博士生来说属于是解放生产力了。

---

• 

  Gerald_Bailey

  —

  看到宣传说60余项评测超过Gemini-3.1-pro和GPT-5.5，先冷静一下，这是发布方自评的科学专业任务榜，日常对话写东西它不一定比得过通用模型，别拿来当ChatGPT用。

---

• 

  DianaBarnes_2021

  —

  作为在读博士，最实用的其实是工具调度台，8000多个科研工具没人能全熟练用，让智能体自动编排工具链这个思路帮我省了太多学习成本。

---

• 

  Amber.PerryX

  —

  全面开源这点必须点赞，官网scienceone.ia.ac.cn直接就能上手，不用申请不用付费，国家队做基础设施就得是这个格局。

---

• 

  KathleenRichardson_2020500

  —

  超算智算速算三位一体这个配置有点东西，尤其速算靠天穹专用硬件跑分子动力学，这种长时程高精度仿真普通GPU真顶不住。

---

• 

  f5kkfztwnkr

  —

  我们组搞材料的，之前看中科院技能大赛里上海硅酸盐所那个吸波材料工具链，输入性能需求模型自己调工具推参数，30分钟顶过去几个月，看得我直接破防。

---

• 

  AnnaTurner_X5

  —

  希望后续院外的开放程度能再大点，现在感觉院内单位能用到的深度和我们这种外面的还是有差距的，别搞成只是名义上开源。

---

• 

  John.Mitchell_2024

  —

  高铁气动载荷那个案例，传统仿真几小时压到秒级，关键参数误差还降了42%，力学所是真下了功夫，工程落地价值比刷榜实在多了。

---

• 

  Edward_Barnes_66

  —

  泼点冷水，第三方独立评测目前太少了，官方口径再漂亮也得等更多外部复现，AI+科学最怕的就是演示很惊艳实际很拉胯。

---

• 

  ik42a_s

  —

  通专一体这个思路我是真认可的，通用模型专业性不够、动不动就一本正经胡说，领域模型又各管一摊互相不通气，磐石想用一个模型把跨学科的数据理解和推理都串起来，方向绝对是对的，但说实话能不能真做到还得看院外普通课题组实际用起来能到多深，别到时候变成院内单位专属。

---

• 

  任然

  —

  问一下有没有搞天文的用过那个光谱智能体，稀有天体识别准确率提升50%这个数听着有点猛，实际部署到郭守敬望远镜上体验咋样？

---

• 

  Jack_Hill_7

  —

  化学那个机器科学家案例是真正的科研闭环，科研人员只要输入创制需求，智能体就自己拆解任务、调度文献挖掘实验设计机器操作计算模拟智算优化五个智能体协同跑，现在已经完成462组样品合成表征和千万级虚拟筛选，前20候选材料里有8个性能超过原来的最优样品，还稳定运行了1500多小时，这才叫AI做科研。

---

• 

  MsBrookeThomas_88

  —

  800万条科学推理数据覆盖200多个任务，据说明年要扩到800个任务，这数据工程量真不是一般团队能扛的，建制化优势体现得淋漓尽致。

---

• 

  JeremyHcward

  —

  说句可能不讨喜的话，磐石的很多亮点本质是模型加平台加专用硬件加海量领域数据的组合拳，文献罗盘也好流体软件也好机器科学家也好，都离不开中科院这套建制化资源和国产算力，脱离了这个体系普通团队想复现同样效果基本没戏，所以开源归开源，能不能跑出这效果是另一回事。

---

• 

  Christopher_Scott_2020

  —

  从1.0到V1.5再到磐石100，一年迭代这么密，中科院这波AI+科学是真上心了。

---

• 

  o0mb6sz1

  —

  磐石2.0英文名ScienceOne Omni起得挺有野心，一个模型通吃数理化天地生，就看这条通专一体的路能走多远了。

---

• 

  DBarnes_X

  —

  它不是给普通人聊天用的，别拿它写文案，人家定位就是科研底座，用错场景纯属浪费。

---

• 

  EthanHall

  —

  谱到分子结构预测能领先领域专用模型这点挺意外的，一般以为通专一体会两头不讨好，结果专业任务上真打过了AlphaFold那类专用模型。

---

• 

  DorisStewart_2020

  —

  青藏科考那个例子印象很深，第二次科考队的科学家专门拿一个青藏专业问题去考它，对比其他通用模型，磐石给的答案更精准更严谨，反复核对之后居然和实际科考情况高度吻合，最后连科考专家都点头认可，这种垂直领域的严谨程度真的是通用大模型给不了的。

---

• 

  vzl36ilns

  —

  证据归因准确率90%这点对科研太重要了，通用模型一本正经胡说的毛病谁受得了，能追溯证据链才敢放心用在正经研究里。

---

• 

  RachelHoward

  —

  三层架构那套科学数据统一编码到领域定向解码的说法听着挺唬人，说白了就是先读懂波谱场这些科学数据再对齐知识最后出答案，能落地就是好架构。

---

• 

  GHallIII

  —

  徐楠那句数据第一架构第二算力第三我太赞同了，科学模态数据不像互联网上的论文新闻随便爬，蛋白质三维结构、晶体结构这些高价值数据得靠长期实验测量才能拿到，还分散在各个数据库和团队手里，得领域专家做质控标注才能用来训练，不是把论文专利下载下来就能训科学基础模型的，这才是真正的瓶颈。

---

• 

  SharonCooper_X130

  —

  免费开源加国产算力适配昇腾海光，自主可控这四个字在当下的分量不用多说了。

No reviews yet