Apache-2.0llmRuns locally14.8B (card); 14,822,046,720 (HF safetensors)
Carbon-X: Specs, License & How to Run
Hugging Face Bio Apache-2.0 14.8B DNA+text chat model with a hybrid 6-mer tokenizer.
- License
- Apache-2.0
- Parameters
- 14.8B (card); 14,822,046,720 (HF safetensors)
- HF likes
- 74 (2026-10-23)
- First seen
- 2026-10-21
TL;DR. Carbon-X is a llm model from Hugging Face Bio (Hugging Face, Zhongguancun Academy, TIGEM / University of Naples Federico II). Parameters: 14.8B (card); 14,822,046,720 (HF safetensors).License: Apache-2.0. It is listed under Closed LLM APIs, ChatGPT.
| Developer | Hugging Face Bio (Hugging Face, Zhongguancun Academy, TIGEM / University of Naples Federico II) |
|---|---|
| Type | llm |
| Architecture | Decoder-only Transformer (HF model_type qwen3; architectures Qwen3ForCausalLM), post-trained from Qwen/Qwen3-14B. Card: 40 layers, hidden size 5120, FFN 17408, GQA (40 query heads, 8 KV heads), SwiGLU, RMSNorm with QK-norm, untied embeddings. Hybrid tokenizer: Qwen3 BPE plus 5,515 DNA tokens (4,096 6-mers, 1,364 tail tokens, plus DNA markers). Vocabulary 157,184. pipeline_tag text-generation. |
| Parameters | 14.8B (card); 14,822,046,720 (HF safetensors) |
| Modalities | text, dna |
| Languages | en |
| Context | 40,960 tokens (card) |
| License | Apache-2.0 — Hugging Face cardData.license is apache-2.0 and Hub tags include license:apache-2.0 on both HuggingFaceBio/Carbon-X and HuggingFaceBio/Carbon-X-GGUF as of 2026-10-23. The README License section says Apache 2.0. A raw fetch of resolve/main/LICENSE 404’d on the BF16 card. Not legal advice. Not a clinical-use grant. |
| Hardware | 1× ~80 GB GPU (BF16, card); Q8_0 GGUF ~16 GB plus KV (GGUF card) — BF16 card: weights fit on a single 80 GB GPU in bfloat16. HF usedStorage is 29,655,940,163 bytes (~27.6 GiB). GGUF card: Q8_0 is 15.8 GB and “needs about 16 GB” with -ngl 99; Q4_K_M is 9.6 GB. Peak VRAM is not independently measured. Tokenizer is custom Python (trust_remote_code=True); Transformers v5 or later (card). (not independently tested) |
| Hugging Face likes | 74 (as of 2026-10-23) |
| Model card | https://huggingface.co/HuggingFaceBio/Carbon-X |
| First seen | 2026-10-21 |
Carbon-X is an open 14.8B decoder-only chat model trained jointly on DNA sequences and natural-language text. Official hosts we opened on 2026-10-23: the ungated card HuggingFaceBio/Carbon-X (created 2026-10-21T10:21:04Z; last modified 2026-10-22T09:55:41Z; 74 likes; Hub downloads field 0 in this check), sibling HuggingFaceBio/Carbon-X-GGUF (created 2026-10-21T11:28:35Z; 12 likes / 2 downloads), and the named base Qwen/Qwen3-14B (Apache-2.0 tag, 513 likes). resolve/main/LICENSE 404’d on the BF16 card.
This identification page hangs off closed LLM APIs and ChatGPT alternatives. That mapping is editorial. Carbon-X is a DNA+text specialist, not a general ChatGPT stand-in. There is no /models/qwen3-14b page on this site, so we did not ship a vs URL. Base-model facts below are copied from those two cards.
Clinical-use limitation (card, copied): Carbon-X “has not been clinically validated, and it should not be used to make medical decisions.” Genetic data is called highly sensitive. Do not treat variant-effect answers as a diagnosis.
What we verified (2026-10-23)
| Field | Value | Source |
|---|---|---|
| Host (BF16) | huggingface.co/HuggingFaceBio/Carbon-X (not gated) | HF API |
| Created / last modified | 2026-10-21T10:21:04Z / 2026-10-22T09:55:41Z | same |
| Likes / downloads | 74 / 0 | same |
| License tag | apache-2.0 | HF cardData |
| LICENSE file fetch | 404 | this check |
| Pipeline tag | text-generation | HF API |
| Base named | Qwen/Qwen3-14B | card base_model |
| Params (card / safetensors) | 14.8B / 14,822,046,720 | card + HF API |
| usedStorage (BF16) | 29,655,940,163 bytes (~27.6 GiB) | HF API |
| Context | 40,960 tokens | card |
| DNA tokenizer | non-overlapping 6-mers + BPE; trust_remote_code=True |
card |
| GGUF sibling | 12 likes / 2 downloads; Q8_0 15.8 GB (card) | HF API + GGUF card |
Clinical and data limits (do not skip)
Copied from the BF16 Limitations section (opened 2026-10-23):
- Not clinically validated. Do not use it to make medical decisions.
- Genetic data is highly sensitive. Local weights, hosted APIs, Spaces, and third-party endpoints can handle inputs differently. The card tells you to understand that path before sending sequences.
- Exact base counts, coordinates, and edits can be wrong. Pair the model with code for those operations.
- It may invent a sequence’s identity or function when the prompt lacks annotations.
- Splice-site and histone-mark GenomeQA rows on the card are 29.8% and 34.6% (self-reported). Long DNA generation is described as out of training intent.
Those sentences stay vendor. forkalt has not run a clinical or bioinformatics eval.
How to run (from official cards, not tested)
Transformers / vLLM (BF16 card). Custom DNA tokenizer: trust_remote_code=True. Transformers v5 or later.
pip install -U transformers
pip install -U vllm
vllm serve HuggingFaceBio/Carbon-X --trust-remote-codeThe card’s Python example loads AutoTokenizer / AutoModelForCausalLM in bfloat16, wraps DNA in a normal chat message, and calls apply_chat_template(..., enable_thinking=False). Sampling defaults copied from generation_config.json on the card: temperature 0.6, top-p 0.95, top-k 20.
GGUF / llama.cpp (GGUF card only): files carbon-x-bf16.gguf (29.7 GB), carbon-x-q8_0.gguf (recommended, 15.8 GB), Q6_K / Q5_K_M / Q4_K_M. Needs llama.cpp b11742 or newer. Example copied from that card:
llama-server -hf HuggingFaceBio/Carbon-X-GGUF:Q8_0 \
-ngl 99 -c 40960 --jinja \
--chat-template-kwargs '{"enable_thinking": false}' \
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0The BF16 card’s agent section still says llama.cpp “is not supported.” We keep that quote next to the GGUF card’s b11742 requirement. Do not mix the two runtimes’ claims.
License
Card tag Apache-2.0 on both siblings. README License section: Apache 2.0. LICENSE blob 404 on the BF16 card (2026-10-23). Not legal advice. Apache-2.0 is not a clinical-clearance statement.
What we did not invent
- DNA benchmark rows (average 66.7%, BioReason Coding 74.9, FrontierVEP 70.4, NextBase 39.4) stay self-reported.
- Cost-per-ClinVar and “1/1,800 of the inference cost” lines stay vendor.
- We do not rank Carbon-X against ChatGPT, GPT-6.1 Sol, or Qwen/Qwen3-14B. Qwen3-14B has no forkalt model page, so there is no vs URL.
Related: closed LLM APIs, ChatGPT alternatives, science/agent VLM Sage-A1.5, serving vLLM. Pixel editors are a different job: PhotoCraft vs GIMP.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| DNA five-benchmark average (card, thinking disabled) | 66.7% (self-reported) | https://huggingface.co/HuggingFaceBio/Carbon-X |
| BioReason Coding (card) | 74.9 (self-reported) | https://huggingface.co/HuggingFaceBio/Carbon-X |
| FrontierVEP (card) | 70.4 (self-reported) | https://huggingface.co/HuggingFaceBio/Carbon-X |
| IFEval prompt-level strict (card) | 79.3% (self-reported) | https://huggingface.co/HuggingFaceBio/Carbon-X |
Deploy options
- local: Hugging Face weights (BF16)
- local: Hugging Face GGUF (llama.cpp b11742+)
- local: vLLM serve
FAQ
Is Carbon-X open weight and Apache-2.0?
The ungated card HuggingFaceBio/Carbon-X (created 2026-10-21T10:21:04Z; 74 likes in the 2026-10-23 API check) tags license apache-2.0. The README License section says Apache 2.0. Sibling HuggingFaceBio/Carbon-X-GGUF (created 2026-10-21T11:28:35Z; 12 likes / 2 downloads) also tags apache-2.0. resolve/main/LICENSE 404’d on the BF16 card in this check. Not legal advice.
Can I use it for clinical or medical decisions?
No sourced clinical validation. The card Limitations section (opened 2026-10-23) says Carbon-X “has not been clinically validated, and it should not be used to make medical decisions.” Genetic inputs are called highly sensitive. That is a vendor limitation, not a forkalt medical opinion.
Does the GGUF work in llama.cpp?
Two official cards, both opened 2026-10-23. HuggingFaceBio/Carbon-X-GGUF says the files need llama.cpp b11742 or newer (HybridDNATokenizer, issue #30825) and that older builds tokenize DNA wrongly. The BF16 card’s agent section still says llama.cpp and other GGUF runtimes are not supported because the DNA tokenizer is custom Python. We copy both. Use the GGUF card for those files. forkalt has not loaded either checkpoint.
Is this a drop-in for Qwen3-14B or ChatGPT?
No sourced bake-off on this site, and there is no Qwen3-14B page here. Carbon-X is a DNA+text chat post-train of Qwen/Qwen3-14B (card base_model). The Carbon-X card’s five-benchmark DNA average (66.7%, thinking disabled) is self-reported, including the “27 points above Qwen3-14B” line. Qwen/Qwen3-14B itself is Apache-2.0, 14,768,307,200 safetensors parameters, 513 likes in the 2026-10-23 API check. We did not create /vs/carbon-x-vs-qwen3-14b.
How do I run it?
Card paths (not tested here): Transformers with trust_remote_code=True, or `vllm serve HuggingFaceBio/Carbon-X --trust-remote-code`. GGUF: llama-server against HuggingFaceBio/Carbon-X-GGUF:Q8_0 with --jinja and enable_thinking false, from llama.cpp b11742+. Sampling defaults on the card: temperature 0.6, top-p 0.95, top-k 20.
Sources & verification
- https://huggingface.co/HuggingFaceBio/Carbon-X
- https://huggingface.co/HuggingFaceBio/Carbon-X-GGUF
- https://huggingface.co/Qwen/Qwen3-14B
- https://github.com/vllm-project/vllm
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.