Language models can already help researchers search literature and synthesize evidence, but scientific work demands grounding and verifiability. In Asta, our agentic platform, users bring complex constraints and treat reports as working artifacts. We built Asta Brief 8B, a small open model for cited report generation, and open-sourced it with training data.
Asta Brief is available in Asta’s Generate a report feature as Fast mode, alongside Claude-powered Thinking mode. It turns research questions and retrieved literature into cited reports, writing in one pass. This reduces generation time by nearly an order of magnitude: Fast mode averages 51.1 seconds per report versus 178.5 seconds for Thinking mode, about 3.5× faster.
We trained Asta Brief from Qwen3-8B, focusing on post-training data and evaluation. It uses tens of thousands of real queries and citation-focused filtering. Open weights allow institutions to run it locally, crucial for sensitive research. We also release an example workflow for adapting reports from PDFs.
The results reflect 2025 comparisons; we haven't rerun against today's frontier models. This work supports broader goals through NSF OMAI, led by Ai2, to build open AI for science. We're sharing our approach to help future models for scientific work.
Source: Hacker News · Summarized by HeadlinesBriefing