MiroMind releases 7B hypothesis-generation model

MiroMind has released MOOSE-Star, a 7B language model post-trained on a curated corpus of 108,000 scientific papers and designed specifically to generate novel, testable research hypotheses. The work was accepted at ICML 2026, and the full collection including model weights and the training dataset is publicly available on Hugging Face. The system is scoped narrowly: it targets hypothesis generation, the step in the scientific method where a researcher proposes a claim that can be tested, rather than literature retrieval or summarization tasks that prior open models have focused on. Training on 108K papers gives the model grounding in existing scientific claims, letting it propose ideas that extend rather than restate the literature. Essentially: MiroMind is offering an open baseline that research teams can fine-tune or benchmark against when building automated ideation pipelines. - The 7B parameter size makes the model runnable on consumer or modest cloud hardware, lowering the barrier for academic labs. - Both weights and dataset are fully open, which is notable given that comparable proprietary systems from larger labs have kept training corpora closed. - ICML 2026 acceptance provides peer-reviewed validation of the methodology, not just a Hugging Face release. Open-sourcing the dataset alongside the weights sets a reproducibility standard that proprietary research-AI tools have largely avoided so far.