Two notable open models arrived on 8 October 2026, aimed at very different jobs. JetBrains, the maker of developer tools such as IntelliJ IDEA, released Mellum2.1, a small model built to act as a coding agent on your own hardware. Hugging Face’s biology team released Carbon-A, a model that reads DNA and predicts where genes that code for proteins are likely to be, along with a large database of its predictions.
Both are “open”: anyone can download the model weights and run them. Both releases are confirmed by the developers. The performance figures in each case are the developers’ own, not independent tests.
Mellum2.1: a small model for coding agents
JetBrains calls Mellum2.1 “the next version of the 12B mixture-of-experts model we open-sourced in June” [1]. A mixture-of-experts model holds many sets of parameters but uses only some of them for each word it produces. Mellum2.1 has 12 billion parameters in total and “2.5B active parameters”, which makes it quick and cheap to run. It is “released under the Apache 2.0 license”, a permissive licence that allows commercial use [1].
The architecture is unchanged from Mellum2. What changed is the training after the initial pre-training stage: JetBrains says it relied mainly on reinforcement learning, in which the model tries tasks and is rewarded for getting them right, using “millions of sandboxed runs across thousands of environments” [1]. JetBrains says the result is a model that “explores a codebase, edits files, and checks its own changes”, which Mellum2 could not do at the level it wanted [1].
JetBrains’ figures (company claims). JetBrains compared Mellum2.1 with Mellum2 and two open models of a similar size, Qwen3.5-9B and Gemma 4 E4B, “using the same evaluation setup for all of them” [1]. It says the biggest gain is in agentic coding. On speed, it says that under heavy load Mellum2.1 “serves almost twice as many tokens as Qwen3.5-9B”, and that a technique called multi-token prediction (MTP) makes a single request “about 1.6 times faster” [1]. No independent benchmark has been published.
What is available now. The weights are on Hugging Face. Builds for popular local tools (GGUF for llama.cpp, Ollama and LM Studio) and the MTP component for the vLLM serving software are “coming soon”, so not yet released [1].
Carbon-A: finding genes in DNA
Carbon-A is “a 1.2B-parameter model that predicts protein-coding regions directly from DNA”, using one model across mammals, plants, fungi and other groups [2]. It reads a “98,304-base-pair context window”, meaning it can consider about 98,000 DNA letters at once [2].
Why it matters: sequencing a genome has become fast and cheap, but working out where the genes are, known as annotation, still lags behind. Hugging Face says “We can now generate genome assemblies far faster than we can understand them” [2].
The database. Hugging Face ran Carbon-A over public GenBank genome assemblies to build the Carbon Annotation Database. The current release covers “48,167 assemblies from 22,617 taxa” with “566 million predicted protein-coding loci” [2]. Those are predictions of where genes may be, not confirmed genes. Hugging Face’s introduction calls them “566 million new gene candidates”; TSN uses the database’s more precise wording.
The evidence (Hugging Face’s own). Across 42 benchmark genomes, Hugging Face reports a macro-averaged nucleotide F1 score of 0.944, a measure of how closely predictions match reference annotations, and says Carbon-A beat the baseline tools it chose to test [2]. Lab work with ActiveSite and UCSD, which read real RNA molecules from cat, Syrian hamster, chicken and Arabidopsis cells, supports some predictions missing from reference annotations. Hugging Face is careful about the limit: these experiments “do not yet establish that those RNAs are translated into proteins—or what those proteins do” [2].
The model weights are published on Hugging Face [3].
What this does not prove
- That Mellum2.1 is faster or better than rivals in practice. The comparisons are JetBrains’ own; no independent test has been published [1].
- That the promised Mellum2.1 builds exist yet. GGUF builds and the MTP head are “coming soon” [1].
- That Carbon-A has found 566 million new genes. These are predicted protein-coding loci, and lab evidence so far supports transcription for some predictions, not protein production [2].
- That Carbon-A beats every gene finder. The 0.944 F1 is Hugging Face’s own benchmark against baselines it selected [2].
The Bottom Line
Mellum2.1 is a compact, permissively licensed model aimed at running coding agents locally, with speed claims that are JetBrains’ own. Carbon-A turns raw DNA into predicted gene locations at very large scale, and Hugging Face is open that its predictions are hypotheses still awaiting lab confirmation. Both are free to download and test, which is the best way to check the claims.
Sources
- Bulat Salimzianov, “Mellum2.1 Gets to Work: A Fast Open Model for Coding Agents”, The JetBrains Blog, 8 October 2026 (company announcement; performance figures are JetBrains’ claims). https://blog.jetbrains.com/ai/2026/10/mellum2-1-gets-to-work-a-fast-open-model-for-coding-agents/
- Hugging Face (HuggingFaceBio), “Carbon-A: Finding genes in known and unknown genomes”, Hugging Face blog, 8 October 2026 (developer announcement; benchmark is Hugging Face’s own). https://huggingface.co/blog/HuggingFaceBio/carbon-annotator-genbank-genome-annotation
- Hugging Face, Carbon-A-1.2B model page, read 9 October 2026. https://huggingface.co/HuggingFaceBio/Carbon-A-1.2B
