Hugging Face's Carbon-X Scores 66.7 on Its Own DNA Suite, a Point Behind GPT-6.1 Sol. Splice Sites Are 29.8, and Two of the Benchmarks Are Ones Hugging Face Built.
Carbon-X is a 14.8-billion-parameter DNA-and-English chat model from Hugging Face Bio, fine-tuned from Qwen3-14B and released under Apache 2.0. On Hugging Face's own evaluation table, with thinking disabled, it averages 66.7% against 67.5% for GPT-6.1 Sol, while GenomeQA splice sites come in at 29.8%. Hugging Face says it built two of the benchmarks in that table, FrontierVEP and NextBase. No independent board has scored it. We read the model card.
By T.W. Ghost
Hugging Face Bio has released Carbon-X, a 14.8-billion-parameter decoder-only model trained jointly on DNA sequences and natural-language text, fine-tuned from Qwen3-14B and published under Apache 2.0. The pitch on the card is a chat model that reads genomes, and the table on the card puts its average across Hugging Face's DNA suite at 66.7%, with thinking disabled, one point behind GPT-6.1 Sol at 67.5%.
Two things about that table decide how much the one point means. First, it is the vendor's table: Hugging Face ran it, Hugging Face chose the tasks, and Sol is the closed comparator on it. Second, Hugging Face says in the same card that it built two of the benchmarks, FrontierVEP and NextBase. A model that lands a point behind a frontier closed model on a suite its own lab partly wrote is a result worth reading, and the column to read first is the weak one: GenomeQA splice sites, 29.8%. We read the model card. Here is what held up.
What You Are Downloading
| Carbon-X | |
|---|---|
| Released by | Hugging Face Bio |
| Size | 14.8B parameters (safetensors about 14.822B) |
| Base | Qwen3-14B |
| Training | DNA sequences and English text, jointly |
| License | Apache 2.0 (card tag) |
| Independent score | None found on any board |
The license is not the catch. Apache 2.0 is on the card tag and is the most permissive license in this month's open-model crop, alongside Recursive Flint. The catch is who graded the exam.
The Benchmarks, All of Them Hugging Face's
| Task (HF card, thinking disabled) | Carbon-X |
|---|---|
| Suite average | 66.7% (GPT-6.1 Sol: 67.5%) |
| GenomeQA organism group | 93.2% |
| GenomeQA motifs | 90.0% |
| GenomeQA enhancer vs promoter | 70.0% |
| FrontierVEP (built by HF) | 70.4% |
| GenomeQA TF binding | 64.0% |
| NextBase (built by HF) | 39.4% |
| GenomeQA histone marks | 34.6% |
| GenomeQA splice sites | 29.8% |
Three honest observations:
- The average hides the spread. Organism group at 93.2% and splice sites at 29.8% are the same model. Anyone using it for splice-site work should read the second number, not the first.
- Two rows are home-built. FrontierVEP and NextBase were created by Hugging Face, per Hugging Face. They may be good benchmarks; they are not external ones, and they should not be described as community boards.
- Sol parity is a same-table claim. 66.7% against 67.5% is one lab's run of both models on tasks that lab selected. It is not an Artificial Analysis score, and none exists for Carbon-X.
The card also carries a throughput cost claim per H200-hour. It is vendor economics we did not confirm, and it stays out.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| 14.8B parameters; Qwen3-14B base; DNA plus English; Apache 2.0 | Verified (Hugging Face model card and tags) |
| Suite average 66.7% vs GPT-6.1 Sol 67.5% | Vendor (HF's table) |
| Per-task scores, splice sites 29.8% included | Vendor (HF's table) |
| FrontierVEP and NextBase built by Hugging Face | Verified as HF's statement (card) |
| Independent DNA board score | Unverified (none found) |
| Per-H200-hour cost claim | Vendor, not confirmed, not printed |
Who Should Use It
Try Carbon-X if you want an Apache 2.0 DNA-literate chat model at a size a single GPU can hold, and you can evaluate it on your own sequences before trusting any row of the card.
Skip the "open DNA champion" framing until a board nobody at Hugging Face wrote scores it, and skip it for splice-site work on the strength of the card alone.
The frontier is elsewhere: on this table the frontier is GPT-6.1 Sol, by a point, on the open model's home turf. Our model comparison has the current standings on the general boards side by side.
Sources
*Not sure whether a domain model beats a frontier general model for your lab's work? Take the free 2-minute quiz and get matched. Then read about the other Apache 2.0 model this month whose only scores are its own: Recursive Flint.*