Hugging Face open-sources Carbon-X, a 14.8B DNA-language model
Open weights make private genomic analysis possible at a fraction of frontier-model cost, which matters for labs handling sensitive sequencing data that cannot leave their infrastructure. The bundled Claude Code and Codex Skills mean bioinformatics teams can plug DNA sequence analysis into existing agent pipelines without building custom integrations. The Qwen3 base also gives practitioners familiar with the Qwen ecosystem a low-friction path to fine-tuning.
Hugging Face released Carbon-X, an open-source 14.8B-parameter model that unifies DNA sequences and natural language in one transformer. Built by continued training of Qwen3-14B with 5,515 added DNA tokens, it was further trained on about 302B tokens, then refined with SFT and reinforcement learning with verifiable rewards. Across five DNA benchmarks it averaged 66.7, close to GPT-6.1 Sol's 67.5 and above Claude Opus 5.5's 60.9, with protein-coding variant effect prediction at 74.9 versus GPT-6.1 Sol's 71. Inference costs roughly 1/1800 of GPT-6.1 Sol per query, and BF16 weights fit on a single 80GB GPU. It answers questions like whether a mutation is benign or pathogenic, and ships with Skills for Claude Code and Codex agent workflows.
发一下前两天Hugging Face开源的一款DNA模型：Carbon-X，在专业基因组任务上逼近GPT-6.1 Sol，单题推理成本约6.1 Sol的1/1800，14.8B
BF16权重可在单张80GB GPU上部署
这是一个让AI直接理解DNA序列，通过自然语言完成基因组分析的模型
比如说，你输入一段DNA序列，就能询问某个突变可能是良性还是致病性、某段序列是否具有基因调控功能，或者预测后续碱基，并文字解释
它核心是把DNA和自然语言统一到了一个语言模型中
基于Qwen3-14B继续训练，扩展了5515个DNA专用Token，混合Tokenizer：文字使用BPE，DNA每6个碱基编码为一个Token
DNA和文字共享同一个Transformer，不需要独立的DNA编码器；模型在Qwen3-14B基础上经约302B token继续训练，再通过SFT与可验证奖励强化学习增强专业能力
在五项DNA基准测试中，其平均得分66.7，接近GPT-6.1 Sol的67.5，高于Claude Opus 5.5的60.9
其中，蛋白质编码区变异效应预测达到74.9，高于GPT-6.1 Sol的71
这对大规模基因组数据库分析、遗传变异研究、生物信息学科研来说是一个低成本、可私有化部署的选择
它还提供了适用于Claude Code、Codex等Coding Agent 的Skill，以及工具调用示例，可以把Carbon-X接入现有 Agent工作流执行DNA序列分析任务
#DNAAI模型 #CarbonX #基因分析模型
