TxGemma

2025-03-25 · Publisher : Google · Domain : Artificial Intelligence

Google releases TxGemma, a collection of open models derived from Gemma 2 for therapeutic development: prediction of molecule properties and conversational discussion, in sizes of 2, 9 and 27 billion parameters.

What changes

Models fine-tuned from Gemma 2 on 7 million training examples, for prediction and conversational analysis of therapeutic data.
Three sizes (2B, 9B, 27B), each with a “predict” version for narrow tasks drawn from Therapeutic Data Commons (for example predicting the toxicity of a molecule).
“Chat” versions in 9B and 27B, able to explain their predictions.
The 27B predict matches or exceeds Tx-LLM, Google's previous generalist model, on 64 of the 66 tasks (45 wins), according to the post.
Agentic-Tx: an agentic therapeutic system powered by Gemini 2.0 Pro and equipped with 18 tools.

### Specifications - Base: Gemma 2. Open models, adaptable to your own data. - Availability: Hugging Face (according to the coverage of the post). ### Pricing Open models, with no usage pricing of their own. ### Benchmarks | Comparison | Result | |---|---| | TxGemma 27B predict vs Tx-LLM | equal or better on 64 tasks out of 66, better on 45 | Scores published by the vendor. ### Key takeaways Specialized models for pharmaceutical research. They are aimed at research teams, not at general-purpose uses.

TxGemma

What changes

Links