Carbon-X trails GPT-6.1 Sol by 2.1 points on the three literature benchmarks, compared with 0.8 across all five tests. Its $300 screen covers uncertain variants whose answers cannot yet be scored.
Hugging Face says its new DNA model, Carbon-X, can read 2.4 million uncertain genetic variants for about $300. The frontier alternative costs between $32,000 and $4 million, by the company's own estimate.
Carbon-X offers credible, low-cost triage while keeping sequences inside the lab. Its claimed parity with GPT-6.1 Sol rests on tests Hugging Face partly built. Its $300 estimate covers variants with no labels against which to score the answers.
A variant of uncertain significance is one with too little evidence to call benign or pathogenic. The ClinVar database holds 2,422,972 of them, and none has a label to score against.
What runs locally
The model card describes a 14.8 billion parameter model built on Qwen3-14B. It had 302 billion more tokens of training on DNA and text, reads DNA in six letter chunks and answers in chat.
Carbon-X runs locally at two sizes, and the full weights fit on one 80 GB GPU. The smaller quantised Q4_K_M file in the GGUF repository is 9.64 GB, the laptop case. The blog shows it running on a MacBook.
The model's own pages disagree on one point. The main card still says llama.cpp is not supported, while the GGUF card says it works.
Comparing the benchmark scores
Hugging Face's parity claim rests on a five test average. The card says Carbon-X averages 66.7% across five DNA benchmarks, close to GPT-6.1 Sol at 67.5%.
Hugging Face built two of those five tests, FrontierVEP and NextBase. It took the other three from the literature, so those three are the cleaner comparison.
On the three literature tests, Carbon-X averages 74.5, GPT-6.1 Sol 76.6 and Opus 5.5 75.8. Carbon-X leads on only one of those three tests.
The gap to GPT-6.1 Sol is therefore 2.1 points, not the 0.8 on the five test average. The chart below sets the two averages side by side for each model.
Carbon-X scores 66.7 on the five benchmark average against 67.5 for GPT-6.1 Sol. On the three literature benchmarks Carbon-X scores 74.5 against 76.6 for GPT-6.1 Sol and 75.8 for Opus 5.5.
View data
| Model | Average of five benchmarks (published) | Average of three literature benchmarks (calculated) |
|---|---|---|
| Carbon-X | 66.7 | 74.5 |
| GPT-6.1 Sol | 67.5 | 76.6 |
| Opus 5.5 | 60.9 | 75.8 |
NextBase also makes Opus 5.5 look weak on the five test average. Opus 5.5 scores 1.1 there at maximum effort, which drags its average to 60.9.
The Hugging Face blog shows that Opus 5.5's score fell from 17.7 at low reasoning effort. Refusals rose from 7.67% to 70.33% over the same range.
Without NextBase, Opus 5.5 averages 75.9 on the other four tests. GPT-6.1 Sol refuses nothing on NextBase, so the collapse is not what separates Carbon-X from it.
Refusals hit a second rival. On FrontierVEP, Gemini 4 Argon's accuracy drops from 58% to 38% as reasoning effort rises, again alongside more refusals, per the blog.
A 2.1 point gap is still a strong result for a model that costs about 1/1,800 as much per problem to run. That supports using it for triage while keeping other models for harder cases.
The comparison may favour the closed models. Each is shown at the effort setting with its best average, and some refuse legitimate prompts.
Refusals are a real cost to anyone using a closed model, and Hugging Face is right to report them. Refusals measure access to an answer rather than DNA skill. The claim under test is parity, and on the three literature tests the gap is 2.1 points.
What the $300 estimate covers
Hugging Face states its cost advantage in two different sizes. The blog's subtitle says 1,000 times cheaper, while its body says 1/1,800 as much per problem.
On the $300 bill, those ratios imply a frontier run of $300,000 or $540,000. Both sit inside the $32,000 to $4 million range, so the two ratios and the range agree.
The $300 is a vendor estimate. NCBI's ClinVar database returns 2,422,972 records classed as uncertain significance, so the count holds. Hugging Face prices measured throughput at $5 per H200 hour.
That buys 60 hours, about 40,000 variants an hour, or $0.000125 a variant. It is a price for inference, meaning running the model, not for clinical interpretation.
The benchmarks score Carbon-X on variants that already have labels, meaning a known answer. So the $300 buys 2.4 million predictions that cannot be checked today.
Uses and accuracy limits
Carbon-X may also have a second role inside larger systems. Hugging Face reports that using it as a specialist helper lifted Kimi K3's DNA index from 49.4% to 60.5%.
The users who gain most are genomics labs and hospital variant teams barred from sending patient data to an outside API. Lewis Tunstall of Hugging Face told WIRED that no data is sent to Hugging Face or any other company.
Such a team has a simple choice. Run Carbon-X locally as a first pass that would rank the variant curation queue for human review, or keep a closed model for the harder cases. The trade is roughly 2 to 3 points of accuracy for control of the data and an inference bill in the hundreds of dollars.
The quantised build costs a little more. On FrontierVEP the GGUF page shows 70.4 for Carbon-X under vLLM and 67.8 for the Q4_K_M build.
The GGUF page lists Q4_K_M as the lowest build on both tests. The 2.6 point drop on FrontierVEP sits against a standard error, a measure of how much a single run's score can vary, of about 1.5 on each single run. The laptop build may cost some accuracy.
The card itself sets the limits. It says the model "has not been clinically validated" and should not be used to make medical decisions. Splice site and histone mark questions remain weak, at 29.8% and 34.6% on GenomeQA.
The card also says the model was not trained to design new sequences. WIRED reports that the developers have not tested it against wet lab experiments, so they cannot say how accurate its predictions are.
WIRED also reports that Carbon-X reads a window of about 50,000 base pairs at a time. Any ranking still needs wet lab confirmation before it touches a patient.
When predictions can begin to be checked
Checking the answers is possible, and the calendar shows when it starts. The NCBI maintenance page says ClinVar archives its monthly release on the first Thursday of each month, and only those releases are archived.
The first archive after launch is 5 November 2026. Anyone who freezes Carbon-X's calls on today's uncertain variants can score the ones later reclassified in the archived releases. Until someone does, the $300 buys predictions with no score.