The lab has open-sourced AstaBrief 8B, a compact model that reads a research question alongside retrieved literature and returns a report carrying citations; it drives the Fast mode in Asta, the lab’s scientific assistant, beside a Claude-driven Thinking mode. Both the weights and the training data are being released.
The goal was to match the report quality of the proprietary models Ai2 had been tracking while trimming cost and delay, and it largely worked. On the full Asta pipeline a Fast-mode report averages 51.1 seconds. Thinking mode needs 178.5 seconds, so Fast mode comes in about 3.5 times quicker.
Training began from Qwen3-8B. The team leaned on supervised fine-tuning and direct preference optimization rather than the reinforcement-learning route used elsewhere, which they found unstable and costly. A redesigned pipeline writes the whole report in one pass instead of section by section.
Grounding was the hardest part. Scientific answers have to stay tied to evidence and must not quietly stretch a study’s conclusions, and researchers need to check the output themselves.
Because the weights are open, institutions can run the model on their own hardware, a requirement when the research touches unpublished or sensitive material. The release ships an example workflow for building reports from a user’s own PDFs.
