01 / The short version
What happened
AstaBrief 8B is an open-weights model designed to generate cited scientific reports from research questions and literature excerpts. Built on Qwen3-8B and trained via SFT and DPO on 47K real-user queries, it generates reports in a single pass, bypassing multi-step pipelines. It averages 51.1 seconds per report in Fast mode versus 178.5 seconds in Thinking mode—3.5× faster. Weights and an example workflow are released to enable local, private report generation.
See the exact references02 / Key takeaways
What you need to know
- 01
AstaBrief 8B is open-sourced with weights and training data for reproducible scientific report generation.
- 02
It uses real user queries (90K filtered to 47K) and proprietary models (Claude, GPT) for training targets.
- 03
Fast mode is 3.5× faster than Thinking mode (51.1s vs 178.5s) with comparable quality.
Keep in perspective
What to watch for
Performance and speed claims are based on 2025-era proprietary models; no comparison to current frontier models has been conducted.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/allenai/astabriefAI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works