据《商业内幕》报道，AI公司Hugging Face周四宣布推出新AI模型Carbon-X，有望帮助科学家更轻松地研究DNA并开发新的基因疗法。Hugging Face称，Carbon-X是一个拥有140亿参数、用于DNA理解和生成的开放模型。
Hugging Face
Carbon-X将DNA序列和自然语言整合到同一个模型中：研究人员可以针对特定DNA序列提出问题、预测突变的影响，并生成新的DNA序列。它还可以充当专业子智能体，帮助其他模型解读它们自身难以理解的DNA序列。
依托千问大模型
Hugging Face高级研究科学家刘易斯·滕斯托尔(Lewis Tunstall)表示，Carbon-X开创了先河，能将DNA分析和生成能力与自然语言相结合，使研究人员能够针对DNA序列提出问题，本质上可以“与DNA对话”。
Hugging Face与意大利和中国的研究实验室合作开发了Carbon-X，这是一个开源模型。Hugging Face通过使用基因组序列和自然语言对现有语言模型千问Qwen3-14B进行训练，开发出了Carbon-X。该模型使用来自人类、其他哺乳动物、植物和真菌的DNA序列进行训练。
Hugging Face表示，在五项DNA理解和生成基准测试中，Carbon-X的表现优于该公司评估的所有其他开源模型，其平均得分接近GPT-6.1 Sol，但每个问题的推理成本仅为后者的1/1800。与Qwen3-14B相比，Carbon-X的得分提高了27分，表明通过直接使用DNA数据进行训练，开放语言模型的性能可以获得提升。
Carbon-X的跑分接近GPT-6.1 Sol
谈到Carbon-X的应用场景时，滕斯托尔举例说，研究人员可以向模型提供一段DNA序列，并询问某种特定突变是否会增加一个人患病的风险。模型随后会给出预测结果并作出解释。它还能够生成新的DNA序列，从而帮助研究人员开发新的基因疗法。
防止过于依赖大公司
滕斯托尔表示，Hugging Face开发Carbon-X的部分动机，是防止生物学研究人员对大型AI实验室产生依赖。他指出，AI已经对数学领域产生了颠覆性影响。
“如果你想真正推动数学研究的前沿发展，你很可能得在OpenAI或Anthropic这样的AI实验室工作，因为它们拥有最好的数学模型。”滕斯托尔称。他认为，其他科学领域也可能出现类似情况。
他补充说：“我们希望为开放研究社区提供工具，让他们不必完全依赖某一家外部供应商。”
研究人员可以在自己的计算机上运行Carbon-X，无需与外部AI公司共享敏感的患者数据。滕斯托尔还表示，在大型模型中观察到的一些行为失准现象，目前尚未在规模较小、专业性更强的模型中出现，例如今年7月OpenAI智能体入侵Hugging Face的知名事件所暴露出的问题。
滕斯托尔表示，通用聊天机器人有时也会拒绝回答研究人员提出的合理生物学问题，因为它们误以为这些问题是在试图“设计生物武器”。他说，为防止模型遭到滥用，Carbon-X的训练数据明确排除了病毒相关数据。
尽管Carbon-X很可能主要供研究人员使用，但滕斯托尔表示，该模型是迈向未来的一小步，AI日后有望帮助提供更个性化的医疗服务。
“我认为，终极目标是实现个性化医疗，也就是我们每个人都能够利用自己的病历和基因组数据，最终找到甚至可能设计出适合自己的治疗方案。”他表示。
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。
According to Business Inside, the launch of the new AI model Carbon-X, announced on Thursday by the AI company Hugging Face, is expected to help scientists study DNA more easily and develop new gene therapy. Hughes Face states that Carbon-X is an open model with 14 billion parameters for DNA understanding and generation.
Hugging Face
Carbon-X integrates DNA sequences and natural languages into the same model: researchers can ask questions about particular DNA sequences, predict the effects of mutations and generate new DNA sequences. It can also act as a professional subdivision of AI agent to help other models interpret DNA sequences that they cannot understand themselves.
LLM.
According to Lewis Tunstall, a senior research scientist in Hugging Face, Carbon-X has pioneered the ability to combine DNA analysis and generation capabilities with natural languages, enabling researchers to ask questions about DNA sequences and essentially “dialogue with DNA”.
Hugh Face, in cooperation with research laboratories in Italy and China, developed Carbon-X, an open source model. Hugging Face developed Carbon-X by training Qwen 3-14B on existing language models using genome sequences and natural languages. The model uses DNA sequences from humans, other mammals, plants and fungi.
Hughes Face states that in the five DNA understanding and generation baseline tests, Carbon-X performed better than all other open source models assessed by the company, with an average score close to GPT-6.1 Sol, but the reasoning cost for each issue was only 1/1800 for the latter. Compared to Qwen3-14B, the Carbon-X score increased by 27 points, indicating that the performance of open language models can be enhanced through direct use of DNA data for training.
Carbon-X runs close to GPS-61 Sol
Turning to the Carbon-X application scene, Tonstor, for example, could provide the model with a DNA sequence and asked whether a particular mutation would increase a person ' s risk of disease. The model then gives projections and explains them. It can also generate new DNA sequences, thereby helping researchers to develop new gene therapy.
Preventing overdependence on large companies
According to Torstol, Hugging Face was partly motivated to develop Carbon-X to prevent biological researchers from becoming dependent on large AI laboratories. He noted that AI had had a destabilizing effect on the mathematics field.
"If you really want to advance the frontier of mathematics, you probably have to work in an AI lab like Open AI or Anthropic because they have the best mathematical models." According to Tonstor. In his view, similar situations could arise in other scientific fields.
He added that “we would like to provide the open research community with the tools to make it unnecessary for them to rely entirely on an external supplier”.
Researchers can run Carbon-X on their own computers without sharing sensitive patient data with external AI companies. Tenstoll also indicated that a number of behavioural irregularities observed in large models have not yet appeared in smaller and more specialized models, such as those revealed by the well-known incident of OpenAIAI agent invasion of Hugging Face in July this year.
Tenstoll stated that general chat robots sometimes refused to answer reasonable biological questions from researchers because they mistakenly believed that they were trying to “design biological weapons”. He said that in order to prevent misuse of the model, the training data of Carbon-X explicitly excluded data related to the virus.
Although Carbon-X is likely to be used primarily by researchers, Tenstoll has indicated that the model is a small step towards the future and AI is expected to help provide more personalized medical services in the future.
“I think the ultimate goal is to achieve individualized medicine, that is, each of us can use our own medical records and genomic data to eventually find, and possibly even design, a cure for ourselves.” He said.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.