Allen Institute for AI released AstaBrief, an 8B open-weights model for cited scientific reports. It runs in Asta’s Fast mode or on local infrastructure. Fast mode averages 51.1 seconds per report. That is 3.5× faster than Asta’s Thinking mode. The model generates full reports in one pass. It uses supervised fine-tuning and direct preference optimization. Training data came from 47K real user queries. Evaluation focused on relevance, coverage, and citation grounding. The model matches the quality of Asta’s proprietary pipeline. Open weights allow institutions to run it on sensitive data. An example workflow for local report generation is also released.
What changed
Asta used proprietary models for report generation in its Thinking mode.
What it unlocks
Generating cited scientific reports in one pass with open weights on local infrastructure.
- 8B parameters
- 51.1 seconds per report in Fast mode
- 178.5 seconds per report in Thinking mode
- 3.5× faster than Thinking mode
What you need to act on it
- Asta’s Fast mode or local infrastructure
Sources