Hugging Face Carbon-X: 14B Open-Source DNA Model Matches GPT-6.1 Sol on Biology Benchmarks at 1/1800th the Cost
Hugging Face’s biology research team released Carbon-X on October 22, 2026, in collaboration with researchers at Zhongguancun Academy in Beijing and the University of Naples Federico II. The model is Apache 2.0, 14.8 billion parameters, and runs on a single 80GB GPU. It is built to answer questions about DNA mutations in natural language.
What Carbon-X Does
A user pastes a DNA reference sequence and a variant sequence. Carbon-X says whether the mutation is likely benign or pathogenic and explains why. That is the core use case: variant interpretation for genetic researchers who want conversational output, not raw numeric predictions.
The practical application is ClinVar screening. ClinVar contains approximately 2.4 million variants of uncertain significance. Hugging Face estimates that screening all of them with Carbon-X costs roughly $300. The same job on GPT-6.1 Sol would cost $32,000 to $4 million depending on context and prompting strategy. For research groups running continuous screens, that cost gap changes what is computationally feasible.
Architecture
Carbon-X starts from Qwen3-14B and adds 5,515 DNA-specific tokens using 6-mers — groups of six DNA base letters that represent recurring structural units. DNA and text share a single token stream, meaning the model can process mixed input without a separate encoding step.
Pretraining continued on 302 billion tokens: 55% DNA sequences from mammals, other vertebrates, invertebrates, plants, and fungi, and 45% English text. The team excluded viral genomes deliberately to reduce dual-use risk. After mid-training, the model went through supervised fine-tuning and reinforcement learning with verifiable rewards — answer correctness checked against known variant classifications.
Input size is capped at approximately 50,000 base pairs per query. The human genome is 6 billion base pairs, so Carbon-X processes fragments, not whole genomes.
Benchmark Numbers
All figures are self-reported unless otherwise noted.
| Model | DNA Benchmark Average (5 tasks) |
|---|---|
| GPT-6.1 Sol | 67.5% |
| Carbon-X (thinking off) | 66.7% |
| Claude Opus 5.5 | 60.9% |
| Kimi K3 | 50.4% |
| Gemini 4 Argon | 49.8% |
| Qwen3-14B (base) | ~39% |
Carbon-X is +27 percentage points over its base model. Two of the five benchmarks used in the average were built by the Hugging Face team; frontier models still lead on non-SNV (single nucleotide variant) tasks, and splice-site and histone modification tasks sit near 30% for all models. The near-tie with GPT-6.1 Sol is a real signal on the tasks measured, not a clean win across all biology.
Tool Use
Kimi K3 rose from 49.4% to 60.5% on the benchmark suite when Carbon-X was available as a tool call — a 11-point lift from calling the specialist model rather than answering from its own parameters. That pattern matters for inference infrastructure: a 14B specialist callable from a larger frontier model is cheaper and more accurate than the frontier model working alone.
Caveats
The model has not been validated against wet-lab experiments. Its variant interpretations are statistical, not clinically tested. The 50K base pair input limit excludes structural variant analysis. The model is weak at long DNA generation despite some early coverage suggesting otherwise.
No independent benchmark runs have been published. The comparison table places Carbon-X above larger closed models on biology tasks, but until third parties verify the numbers, the gap should be treated as a claim.
Significance
Specialist open models matched to domain-specific benchmarks at orders-of-magnitude cost reduction is the direction genomics research is moving. AlphaGenome (Google DeepMind) predicts DNA effects numerically. Evo 2 (Arc Institute/NVIDIA) generates full genomes. Carbon-X sits in the conversational layer — a chatbot for DNA that researchers can deploy privately on their own hardware. The privacy case is specific: hospitals and research hospitals face data governance constraints that prevent uploading patient-adjacent sequences to frontier model APIs. A locally hosted 80GB checkpoint removes that constraint.