完全免费，Claude Science开源平替，DeepSeek/GLM想用哪个用哪个

marsbitPublished on 2026-07-07Last updated on 2026-07-07

Abstract

近日，开源社区推出了“Claude Science”的平替项目OpenScience，引发科研人员关注。Claude Science是Anthropic公司推出的AI科研工作台，集成了文献检索、假设生成、代码实验到论文撰写等全流程工具，并内置大量科学数据库和多智能体系统，可显著提升科研效率。但其仅支持macOS/Linux、需付费且仅限Claude模型使用，门槛较高。 OpenScience由YC孵化的Synthetic Sciences团队开发，同样提供全流程AI科研支持，但采用“模型不可知”设计，用户可自由选用DeepSeek、GLM、Claude、GPT等国内外模型，甚至本地部署。项目开源且安装简便，内置技能包更丰富。团队强调其独立性，与Anthropic无关。该开源方案降低了使用门槛，为科研工作者提供了灵活、开放的AI协作工具选择。

学术人狂喜!!

Anthropic的Claude Science上线还不到一周,开源社区就亮出了自己的答卷。

一家由YC孵化的AI科研团队,交出了「Claude Science开源平替版」OpenScience。

同样是覆盖文献检索、假设生成、代码实验到论文撰写的全流程AI科研工作台,但它不绑定任何一家模型厂商。

DeepSeek、GLM、Claude、GPT......甭管国内还是国外,你想用哪个就用哪个。

而且项目采用对开发者最为友好的Apache 2.0协议,仅需一行命令即可安装。

消息一出,项目直接冲上X热搜。人们纷纷直言:

这才是scientific AI该有的样子。(A社:报我名得了)

Claude Science虽强,但用不了啊...

大约5天前,Anthropic在MIT Technology Review的一场闭门活动上,正式发布了Claude Science。

这是一个专门面向科学家的AI工作平台,提供研究人员最常用的各种工具和软件包。

举个例子,以前一个科研人员要完成一项研究,得在PubMed上查文献,用Jupyter写代码,拿R跑统计,SSH连集群提交任务,再用各种工具画图、写论文。

十几个窗口来回切,光是工具间的「转换」就足够消耗大量精力。

而Claude Science想干的,就是把这些全塞进同一个工作台。

具体来看,它做了几件比较关键的集成:

数据库和工具链层面,内置了60多个科学数据库连接器和预配置技能包,覆盖基因组学、单细胞分析、蛋白质组学、结构生物学、化学信息学等常见科研领域。

你用自然语言提问,专业Agent会自动跨库查询,UniProt、PDB、Ensembl、ChEMBL、GEO这些数据库不用挨个翻。

它还接入了英伟达的BioNeMo Agent Toolkit,可以直连Evo 2、Boltz-2、OpenFold3等生命科学模型。

执行层面,它引入了多智能体架构。

主Agent负责整体规划,子Agent并行处理不同任务,还有一个Reviewer Agent,专门负责做事实核查,比如检查引用、验证计算结果、标注潜在错误。

生成的结果也不只是文本输出,3D蛋白质结构、基因组浏览器轨道、化学结构式这些内容,它都能原生渲染。

而且每张图表都会同时保留生成代码、运行环境、自然语言说明和完整对话历史。

在某些场景里,科学家甚至可以直接用一句话改图,系统自动回写底层代码。

算力层面,Claude Science可以直接对接你实验室已有的基础设施。

笔记本、Linux服务器、HPC集群登录节点都行,通过SSH连接或者Modal账户按需调用云端GPU,从单卡扩展到数百卡。

大规模数据集只需加载一次,敏感数据不用离开你自己的系统,只有每一步分析需要的上下文才会发送给Claude。

早期的内测用户已经跑出了一些实际案例。

艾伦研究所的神经科学家Jérôme Lecoq用它搭了一个多智能体「计算评审模板」,包含约20个自定义技能,让子Agent读数千篇论文、提取核心观点和定量数据,再逐章生成综述。

就这么说吧,以前写一篇综述要两年,现在他手上已经有大约10篇——

许多超过100页,且引用全部经过Reviewer Agent核验。

而UCSF脑肿瘤中心的Stephen Francis用它做胶质瘤的分子流行病学研究,跑germline变异分析。

他说Claude Science把原本需要的时间压缩到了十分之一,而且他的团队独立验证了结果,确认分析既快又靠谱。

结合今年3月哈佛物理学家马修·施瓦茨对AI科研能力的评价,目前Claude的水平大约相当于一位研二学生。

他在Anthropic官方博客发表了一篇客座文章「Vibe Physics: The AI Grad Student」,里面记录了他用Claude Opus 4.5完成一篇理论物理论文的全过程。

当时他给出的结论是:

当前AI的科研能力大约相当于一名研二学生,能干活、不喊累,但每一步都需要导师盯着。

这个判断后来也被Anthropic写进了Claude Science的技术文档,作为产品定位的一个校准点。

不过,Claude Science目前有几个硬限制:

只支持macOS和Linux

仅面向Pro/Max/Team/Enterprise付费用户

在平台上只能用Claude自家模型

这几道门槛叠在一起,尤其对国内科研团队来说,Claude Science就成了「可望不可即」的东西。

好消息:开源平替来了

瞄准以上限制,开源项目OpenScience应运而生。

其背后团队名为Synthetic Sciences,2025年在旧金山成立,今年刚从YC 2026冬季批次毕业。

创始团队的野心不小,要造一个平台,让科学家把复杂的研究任务直接委托给「AI联合科学家」(AI co-scientists),从文献综述到假设生成到实验执行到论文撰写,整条链路都让AI自主跑通。

他们内部有一个核心判断:

科学基础模型需要具备真正的「研究品味」(research taste),而这种品味靠单纯堆参数堆不出来,必须产品和模型两条腿走路,用产品收集高质量的科研过程数据,再用这些数据训练出有品味的模型。

OpenScience,就是这条路线落地的第一个产品。

虽然OpenScience的使命和Claude Science一样,但它们有一个根本性区别:

模型不可知(model-agnostic)。

用Synthetic Sciences自己的话说:

科学AI应该是开放的,不该由一家公司垄断人类用来探索发现的工具,更不该由它来决定谁有资格使用。

所以在这个平台,Anthropic、OpenAI、Google、DeepSeek、GLM......只要你手里有API Key,都能直接接。

甚至可以跑本地模型,用Ollama部署,数据一个字节都不用出你的机器。

你的Key留在本地,请求直连模型提供商,不经过任何中间服务器。

而且,OpenScience支持按请求切模型。

同一个工作台里,你可以这一步用Claude,下一步换DeepSeek,不需要改任何配置。

功能层面,OpenScience甚至比Claude Science更激进——

内置250多个研究技能包,是Claude Science的4倍多,覆盖ML、计算生物学、化学信息学等方向,而且全部可读、可编辑、可扩展。

安装也很简单,终端一行命令:

即开即用,浏览器自动弹出工作台。首次运行选个模型来源,填上API Key,就能开始干活。

想全局安装也行:

嫌配Key麻烦的话,团队还提供了一个托管平台Atlas——

充值钱包直接调用多家前沿模型,不用挨个配Key,还有持久化的研究图谱和云端算力。

但这个Atlas不是必须的,你用自己的Key跑OpenScience,一样可以完全免费用起来,没有门槛。

One More Thing

有意思的是,翻到OpenScience的GitHub页面最底部,你会看到一段特意加上的声明:

OpenScience is an independent project. It is not affiliated with, endorsed by, or sponsored by Anthropic. “Claude” is a trademark of Anthropic, PBC, used here only to describe compatibility.

翻译翻译,我们是独立项目,跟Anthropic没有任何关系。提到「Claude」纯粹是说兼容性,别多想。

看来当初的「龙虾」给整个开源社区留下的印象,还是太深刻了。

OpenClaw几度更名在前,OpenScience这次直接把撇清关系的声明焊死在README第一版里。

无他,先活下来,再谈平替(doge)。

开源地址:

https://x.com/i/trending/2073904804829741364?s=20

参考链接:

[1]https://x.com/SynScience/status/2073829478393086311?s=20

[2]https://x.com/i/trending/2073904804829741364?s=20

[3]https://www.openscience.sh/

[4]https://www.anthropic.com/news/claude-science-ai-workbench

本文来自微信公众号“量子位”,作者:一水

Trending Cryptos

CitreaCTR

wrapped stUSDTWSTUSDT

Velodrome FinanceVELODROME

BrevisBREV

PancakeSwapCAKE

Earning $7.2 Million in One Week: How Does pump.fun Actually Make Money?

Pump.fun, a prominent Solana-based meme coin launch platform, recently disclosed a weekly protocol revenue of $7.2 million. The income primarily stems from its core businesses: the Bonding Curve mechanism, which facilitated $553 million in weekly volume, and PumpSwap, which saw $1.65 billion. The platform is allocating 50% of its net protocol revenue to buy back and burn its native PUMP token, having already burned 41.8% of the circulating supply with $3.7 million repurchased in the past week. This establishes a direct link between platform performance and token value. Beyond strong financials, Pump.fun is evolving from a simple launchpad into a broader ecosystem. Key initiatives include optimizing the user experience with faster swap times (300-400ms) and easier fiat on-ramps, which boosted daily deposit volume by 21%. It's also enhancing developer tools via Terminal and fostering community engagement through a bounty system that has distributed over $600,000 in rewards. The report signals Pump.fun's transition towards a sustainable Web3 business with diversified revenue streams, a value-accrual model for its token, and a focus on product maturity and ecosystem growth over mere meme coin hype.

marsbit40m ago

Earning $7.2 Million in One Week: How Does pump.fun Actually Make Money?

marsbit40m ago

Tiger Research: Three Strategies for Financial Institutions to Keep Up with the Tokenization Wave

Tiger Research's in-depth report analyzes the strategic choices for financial institutions entering the rapidly growing Real-World Asset (RWA) tokenization market. With the market projected to be worth $25-36 billion and a lack of complete regulatory frameworks in many jurisdictions, institutions face three options: waiting for domestic legislation, using regulatory sandboxes for limited experiments, or moving first into established overseas markets to gain a competitive edge. Tokenization is not magic; it requires meticulous preparation. Before entry, institutions must strategically plan across six core areas: choosing a jurisdiction, obtaining necessary licenses, defining the asset type, targeting the appropriate investor base, selecting settlement currencies, and designing operational requirements like custody and governance. The report outlines two main operational paths. The first is a direct jurisdictional path, establishing a legal presence in mature markets like Hong Kong, Singapore, or the U.S., often leveraging local licensed platforms to accelerate market entry. The second is a chain-native path, using platforms like Ondo or Plume that are built with regulatory compliance embedded, allowing for faster, more flexible market access without being tied to a specific jurisdiction. Ultimately, the report advises against waiting for perfect regulation. The preparation process can take 6-12 months and requires thorough legal review. Using the example of a mid-sized securities firm, it details steps from evaluating existing entities to final execution. The core message is that accumulating real-world operational experience is paramount, as the market is moving ahead and will not wait for latecomers.

Foresight News1h ago

Tiger Research: Three Strategies for Financial Institutions to Keep Up with the Tokenization Wave

Foresight News1h ago

‘Incentivize democracy’ at the cost of security? Inside Bonk’s $20M exploit

BonkDAO, the governance body for the memecoin BONK, suffered a $20 million exploit due to a malicious governance proposal. The attacker acquired enough BONK to create a proposal with a 1% vote threshold, which then authorized transferring 4.4 trillion BONK to their wallet. Security analysts criticized the low approval threshold and lack of safeguards like lock periods for treasury transfers, calling it an operational security failure. The incident fueled debate about balancing decentralized governance with security. Following the news, BONK's price dropped 10%, though on-chain data indicated continued accumulation by large holders since June. The project is collaborating with law enforcement and exchanges to mitigate the situation.

ambcrypto2h ago

‘Incentivize democracy’ at the cost of security? Inside Bonk’s $20M exploit

ambcrypto2h ago

Zoomex X Space Recap With David James and the World Cup Trading Panel

Zoomex hosted a World Cup-themed X Space featuring England goalkeeper David James and a panel of traders. The discussion, part of a charity pledge, drew parallels between goalkeeping and trading, focusing on preparation, psychology, and execution. James emphasized that pressure for a goalkeeper comes during periods of inactivity, not sustained attack, and highlighted the importance of concentration. He described two modes for penalty saves: meticulous preparation using data on shooters, and instinct, which is improved by information—a concept traders directly related to market analysis. He discussed how elite preparation turns instinct into trained execution, making spectacular saves feel like practiced outcomes rather than miracles. Errors are now treated as data for correction, not lingering ghosts. James analyzed tournament threats, noting France's exceptional speed across the pitch and suggesting that teams like Mexico and Spain, yet to concede, face a psychological test. On England's chances, James was unequivocal, predicting they would win and nominating the UEFA Foundation for the associated charity donation. The session consistently linked thorough preparation in both football and trading to controlling emotions and executing under pressure, with the core lesson being that preparation eliminates nervousness by replacing unknowns with a plan.

TheNewsCrypto2h ago

Zoomex X Space Recap With David James and the World Cup Trading Panel

TheNewsCrypto2h ago

Unitree's IPO Frenzy: The Real Mystery is How It Will Spend the 42 Billion Raised

Unitree, a Chinese robotics company, is set for a public listing after its IPO registration was approved by regulators. The company, which started with quadruped robots and has expanded into humanoids, plans to raise approximately 4.2 billion yuan through its offering. The article traces Unitree's rapid growth from its founding in 2016 to its current status. It highlights key milestones like the 2021 CCTV Spring Festival Gala performance, the 2023 launch of its affordable Go2 robot dog and the H1 humanoid robot, and a series of subsequent product launches. By 2025, the company reported revenue of 1.71 billion yuan, profitability, and sales exceeding 5,500 humanoid robots. As the first publicly-listed humanoid robot company on China's STAR Market, Unitree's main challenges are sustaining growth and deploying its newly raised capital effectively. The humanoid robot sector in China is crowded, with over 140 companies. Competitors include UBTech (focusing on industrial and consumer markets), Fourier, and international players like Tesla Optimus and 1X NEO. The article outlines three critical challenges for Unitree: establishing a strong second product line beyond its quadruped robots, maintaining its price advantage while ensuring quality, and successfully advancing its embodied AI capabilities through partnerships like the one with NVIDIA for the H2 Plus platform. Unitree's likely strategy involves a "developer tools + industry benchmarks" approach: using low-cost models like the R1 and G1 to build developer adoption and volume, leveraging high-end platforms for AI training, and securing pilot projects in sectors like logistics and manufacturing to build case studies. The company's future success hinges on converting its current momentum in shipments and pilot programs into sustainable, large-scale commercial contracts as the broader market evolves.

marsbit2h ago

Unitree's IPO Frenzy: The Real Mystery is How It Will Spend the 42 Billion Raised

marsbit2h ago

Trading

Spot

Hot Articles

How to Buy GLM

Welcome to HTX.com! We've made purchasing Golem (GLM) simple and convenient. Follow our step-by-step guide to embark on your crypto journey.Step 1: Create Your HTX AccountUse your email or phone number to sign up for a free account on HTX. Experience a hassle-free registration journey and unlock all features.Get My AccountStep 2: Go to Buy Crypto and Choose Your Payment MethodCredit/Debit Card: Use your Visa or Mastercard to buy Golem (GLM) instantly.Balance: Use funds from your HTX account balance to trade seamlessly.Third Parties: We've added popular payment methods such as Google Pay and Apple Pay to enhance convenience.P2P: Trade directly with other users on HTX.Over-the-Counter (OTC): We offer tailor-made services and competitive exchange rates for traders.Step 3: Store Your Golem (GLM)After purchasing your Golem (GLM), store it in your HTX account. Alternatively, you can send it elsewhere via blockchain transfer or use it to trade other cryptocurrencies.Step 4: Trade Golem (GLM)Easily trade Golem (GLM) on HTX's spot market. Simply access your account, select your trading pair, execute your trades, and monitor in real-time. We offer a user-friendly experience for both beginners and seasoned traders.

1.9k Total ViewsPublished 2024.03.29Updated 2026.06.02

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of GLM (GLM) are presented below.