# Carbon-X (slim)

> Carbon-X is Hugging Face's open-weight 14.8B language model for DNA, a fine-tune of Qwen3-14B that reads raw DNA in a chat and answers in English about variant effects and regulatory regions, with a 40,960-token context, under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/carbon-x.md (~11,200 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/carbon-x.json · canonical https://www.anchorterminal.com/tools/carbon-x
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-23 · pricing https://www.anchorterminal.com/tools/carbon-x#pricing · reviews https://www.anchorterminal.com/tools/carbon-x#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #38 of 107 open-weight models · #30 in Open-weight models · not agent-ready · confidence low**


Assessment: A DNA specialist cousin that another agent calls for variant-effect and sequence questions, with DNA kept on the machine. Hugging Face's Q4_K_M GGUF needs about 12 GB and Q8_0 about 19 GB, on llama.cpp b11742 or later. It isn't clinically validated, it isn't trained for tool calls or reasoning, and the card contradicts its own GGUF repository on llama.cpp support.

## Facts

- Kind: Open-weight model · vendor: Hugging Face (HuggingFaceBio) · category: Open-weight models · legal entity: Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub) · provenance 53/100
- Packages: huggingface `HuggingFaceBio/Carbon-X`, huggingface `HuggingFaceBio/Carbon-X-GGUF`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/HuggingFaceBio/Carbon-X and HuggingFaceBio/Carbon-X-GGUF, read from both cards, the files and their metadata, the chat template, the custom tokenizer code, the launch post (a Space), the huggingface/carbon GitHub repository and llama.cpp's own code on 23 October 2026. Nothing was downloaded or run. Hugging Face has no hosted API for it
- Model: 14,822,046,720 parameters (BF16), Qwen3ForCausalLM with 40 layers, hidden size 5120, 40 query heads and 8 KV heads, SwiGLU and QK-norm, untied embeddings. Qwen3-14B with 5,515 DNA tokens added (4,096 6-mers, 1,364 tail tokens and three markers), mid-trained on 302B tokens, then fine-tuned and trained with GRPO on DNA tasks with thinking off. Text and DNA in, text and DNA out
- Files: Two BF16 safetensors shards (29.64 GB) with chat_template.jinja, tokenizer.json, a custom tokenizer.py, dna_config.json, two agent examples and a carbon-dna skill for coding agents. The GGUF repository holds BF16 (29.7 GB), Q8_0 (15.8 GB), Q6_K (12.6 GB), Q5_K_M (11.1 GB) and Q4_K_M (9.6 GB). No pickle files. Hugging Face lists a SHA-256 for every LFS file
- Licence: Apache 2.0 in the card's metadata and its License section, with no LICENSE file in either repository. The base Qwen/Qwen3-14B is Apache 2.0 with a LICENSE file, and the tokenizer builds on Qwen/Qwen3-4B-Base, also Apache 2.0. No use policy
- Runtimes: Transformers v5 and vLLM per the card, both with trust_remote_code for the DNA tokenizer. llama.cpp from release b11742 with the publisher's GGUF files. SGLang has the qwen3 architecture in its released code but nobody documents it for Carbon-X. No Ollama library entry and no LM Studio catalogue page
- Context: 40,960 tokens in config.json, the card and the GGUF metadata, rope_type default. Mid-training used 8,192-token sequences and post-training examples were under 16K tokens, and the card says longer inputs weren't evaluated. tokenizer_config.json says 131,072 as model_max_length
- Tools and reasoning: The template is a training variant of Qwen3's, with tools as JSON inside <tools> and calls as JSON inside <tool_call>, and it keeps a think block in every assistant turn. The card doesn't describe the model calling tools; it documents Carbon-X as a tool another agent calls, with an OpenAI-style tool example, an Inspect example and a skill. Post-trained and evaluated with enable_thinking false, which the card recommends
- Hardware: The card says the BF16 weights fit one 80 GB GPU and serve with vLLM on a single GPU. The GGUF card says Q8_0 needs about 16 GB plus about 0.16 MB of KV cache per token
- Training data: Mid-training mix with shares of 302B tokens, raw DNA 55% (eukaryotic genomes and OpenGenome2), English 20% (FineWeb-Edu, DCLM, FinePDFs-Edu), biology text 15% (FinePDFs, PubMed, bioRxiv, FineWeb, Biostars and synthetic RefSeq descriptions) and unnamed instruction data 10%. The SFT mix of 20 subsets and the RL prompt counts are described without dataset names. Eukaryote-infecting viral genomes were excluded, per the launch post. No knowledge cut-off

- Scores: Capability pending, Reach 90, Risk 65, Agent interface 35, Documentation and upkeep 77 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, the publisher's Q4_K_M GGUF is 9.6 GB, or 11.6 GB with a fifth added. · Risk, File safety 15, safetensors and GGUF only, but the documented loading path runs the repository's tokenizer.py through trust_remote_code. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 17, everything but a knowledge cut-off.
- Sources: 23, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/carbon-x.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/carbon-x.svg` or a link to https://www.anchorterminal.com/tools/carbon-x from a page under huggingface.co/huggingfacebio, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest start is `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 0`, on llama.cpp b11742 or later
2. With vLLM or Transformers pass trust_remote_code, and expect the tokenizer to fetch Qwen/Qwen3-4B-Base at a pinned revision on first load
3. Keep enable_thinking false and the card's sampling (temperature 0.6, top_p 0.95, top_k 20), the settings every reported result used
4. Keep DNA in files and pass sequences by name, compute positions and base changes in code and send them as context, the card says the model miscounts long sequences
5. Only A, C, G and T form 6-mers, so filter N and other IUPAC codes first or they become a single oov token
6. Strip the dna tags from answers if you need bare sequence, llama.cpp keeps them in the text
7. Cap max_tokens for next-base questions, long DNA generation is slow and degrades

## Connect

```bash
# llama.cpp b11742 or later (older builds tokenize DNA wrongly):
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 0 --port 8080
# or, with vLLM:
vllm serve HuggingFaceBio/Carbon-X --trust-remote-code
```

```bash
curl http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"messages": [{"role": "user", "content": "Is this sequence more likely coding or non-coding? ATGGCGCTGTGGATGCGCCTCCTG"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 3/5, desk reviews from public material, no calls made)

- ★★★☆☆ No repository code on the GGUF route, no clinical validation (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Squashed history and a card at odds with its GGUFs (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Built to be called, not to call tools (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Apache 2.0 and a 9.64 GB Q4_K_M from Hugging Face (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-23; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 2 to 4. Every reviewer notes that the card says Carbon-X isn't clinically validated and that no wet-lab validation is reported, and three note that its DNA benchmark figures are Hugging Face's own. Read it as an Apache 2.0 DNA model with small publisher GGUFs, a card that contradicts its own GGUF repository on llama.cpp, and no trained tool calling.