Carbon-X Open-source DNA The model is reportedly based on Qwen3-14B training

Article

Hugging Face uses an older version to Qwen3 train DNA model: 14B Benchmark score approaches GPT-6.1 Beating AI reports that Hugging Face, together with Beijing Zhongguancun Institute and an Italian research institution, has open-sourced a biological model Carbon-X, enabling researchers to analyze DNA as easily as using ChatGPT. For example, by entering segments of normal and mutated DNA and directly asking, “Could this change cause disease?”, the model can provide predictions and explanations. It can also identify regions affecting gene activity, determine DNA sources, and even complete gene sequences.Carbon-X is not based on the latest model, but rather on 2025 released in Qwen3-14B. The team added a vocabulary specifically for recognizing DNA, encoding every 6 DNA bases as Token, and then continued training with approximately 302000000000 Token of DNA, biological text, and instruction data, enabling the model to both understand gene sequences and answer questions in natural language.

In the officially released 5 DNA benchmark tests,Carbon-X achieved an average score of 66.7%, approaching GPT-6.1 Sol’s 67.5%, and surpassing all open-source models tested. In particular, it outperformed all tested models in mutation assessment in protein-coding regions and DNA sequence-completion tests. According to the team’s estimate of computing costs, its inference cost per question is approximately GPT-6.1 Sol API of 1/1800.

The model has approximately 14800000000 actual parameters, and its weights and locally quantized version have been released, allowing researchers to process sensitive genetic data on their own devices.

Original link https://m.theblockbeats.info/flash/373214

Key points

Beating AI reports that Hugging Face, together with Beijing Zhongguancun Institute and an Italian research institution, has open-sourced a biological model Carbon-X。
According to the report,Carbon-X can compare normal and mutated DNA and predict whether a change may cause disease; it can also identify regions affecting gene activity, determine DNA sources, and complete gene sequences.
Carbon-X uses 2025 released in Qwen3-14B as its base model, and adds a vocabulary specifically for recognizing DNA.
The team encodes every 6 DNA bases as Token, and continues training using approximately 302000000000 Token of DNA, biological text, and instruction data.
In the officially released 5 DNA benchmark tests,Carbon-X achieved an average score of 66.7%，GPT-6.1 Sol scored 67.5%。
According to the report,Carbon-X outperformed all tested models in mutation assessment in protein-coding regions and DNA sequence-completion tests.
According to the team’s estimate of computing costs,Carbon-X has an inference cost per question of approximately GPT-6.1 Sol API of 1/1800。
According to the report,Carbon-X has approximately 14800000000 actual parameters, and its weights and locally quantized version have been released.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Carbon-X is an Qwen3-14B analysis model further trained on the general-purpose language model DNA. It converts DNA sequences into Token that the model can process, allowing it to both read gene sequences and provide predictions and explanations in natural language.

Why it matters to readers

If the open weights and locally quantized version mentioned in the report are indeed available, researchers can analyze DNA on their own devices and process sensitive genetic data locally.

Beginners can understand it as an open-source model that can “read DNA and answer questions,” but the reliability of its disease predictions cannot be determined solely from the benchmark scores reported.

Risks and unknowns

The benchmark scores were officially released; the evidence contains no independent replication results.
The per-question cost ratio is based on the team’s estimate; the specific calculation method and conditions of use are not stated.
The evidence does not provide direct links to the model weights, quantized version, or official release page.
The evidence does not specify the model license, deployment hardware requirements, or specific safeguards for processing sensitive genetic data.
The evidence does not specify whether disease prediction can be used for clinical decision-making.

Carbon-X Open-source DNA The model is reportedly based on Qwen3-14B training

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center