In 2026, Hugging Face released Carbon, a family of open genomic models. The headline claim is that Carbon-3B matches Evo 2, the leading open DNA model, while running far faster. Around the same time, a separate project, OpenMed, released a clinical AI agent that works in the terminal. Both are open source under the Apache 2.0 license.

We read Carbon's own model card to see what "matches" means. The answer is more specific than the headline.

Does Carbon match Evo 2?

The Carbon-3B model card compares it with Evo2-7B on six zero-shot tasks. We took the scores as published and subtracted.

Source: Carbon-3B model card, Hugging Face. Scores are those published by the model authors, not independently reproduced. Zero-shot evaluation.

View data

task	gap
Sequence recovery (eukaryote)	1.68
ClinVar non-coding variants	1.35
BRCA2 variant effect (AUROC)	1.11
ClinVar coding variants	-0.44
Synonymous codon discrimination	-2.7
Triplet-expansion discrimination	-3.23

On sequence recovery in eukaryotes Carbon scores 61.54% against 59.86% for Evo2-7B. On BRCA2 variant effect it scores 84.63% against 83.52%. On non-coding ClinVar variants it is ahead 91.14% to 89.79%. It is behind on coding ClinVar variants (92.89% against 93.33%), on synonymous codon discrimination (88.89% against 91.59%) and most clearly on triplet-expansion discrimination (85.20% against 88.43%).

Averaged over the six tasks, Carbon-3B is 0.37 points behind. "Matches Evo 2" is fair as a summary of a close contest. It is not a clean win. The model card also reports that Carbon matches Evo2-7B on a long-context retrieval test at 393,000 bases, with scores of 0.79 to 0.80.

Where does the speed come from?

The speed claim varies by source. The model card says Carbon generates sequence 150 times faster than the Evo 2 family and produces over 100,000 bases per second on a single H100 GPU. Other write-ups quote 250 times. We could not reproduce any of these, so we treat them as the authors' claims.

The mechanism is the tokenizer, and it shows up in the training numbers.

Tokens seen in training (trillions)DNA bases seen in training (trillions)

Source: Carbon-3B model card; Evo 2 paper (bioRxiv, 2025) and Arc Institute. Carbon: 1 trillion 6-mer tokens, about 6 trillion bases. Evo 2 7B: 2.4 trillion tokens; Evo 2 40B: 9.3 trillion tokens, at single-base resolution.

View data

model	tokens	bases
Carbon-3B	1	6
Evo 2 7B	2.4	2.4
Evo 2 40B	9.3	9.3

Evo 2 reads one base per token, at single-base resolution. Carbon groups DNA into 6-mers, so each token holds six bases. Each generated token covers six bases instead of one, which is a plausible source of a large speed gain.

It comes with a trade-off. Carbon's native context is 32,768 tokens, about 197,000 bases, extended to about 393,000 bases at inference. Evo 2 supports a context of one million bases. Carbon trades length for speed.

What is OpenMed?

OpenMed Agent is a different kind of project. It is a terminal-based runtime for clinical workflows, in public preview since May 2026 and at version 0.1.2 on June 7, 2026. According to its documentation, it ships with 76 native tools and 15 deterministic workflows covering tasks such as prior authorization, coding audits and discharge handoffs. Nothing is finalized without a draft and an approval step, and every tool call is recorded with its provenance. Sessions stay on the user's machine while medical processing runs on hosted services during the preview. It reports no built-in analytics.

The two projects share a design choice. Both make the working visible: Carbon publishes its scores and tokenizer, and OpenMed shows each tool call and requires review before output is final.

Limits

We relied on published model cards and documentation. We did not run either project. Benchmark scores come from the model authors. The Evo 2 token counts come from published summaries of the paper, and the speedup varies between 150 and 275 times depending on the source. OpenMed is a preview and its feature counts may change.

Sources

1. Hugging Face, Carbon-3B model card and Carbon repository: benchmark scores, tokens, context length, throughput, license.
2. Arc Institute, Evo 2 repository and paper on bioRxiv: model sizes and training tokens.
3. Hugging Face releases Carbon and Carbon-3B: a 3B DNA model that matches Evo2-7B at 150x the speed: other speedup figures quoted.
4. Mervin Praison, OpenMed Agent preview and OpenMed.