Introducing AstaBrief 8B, an open model that turns complex research questions + literature excerpts into cited reports. Run it locally on your own hardware, with open weights + training data you can inspect & build on. 🧵 🤗 Download:

AstaBrief powers Fast mode in Asta’s Generate a report feature. It writes reports in one pass, skipping intermediate steps that summarize & group the retrieved excerpts so researchers can get a preliminary report sooner. 🌐 Try Fast mode:

We built AstaBrief on Qwen3-8B using queries from ScholarQA & our earlier literature-synthesis systems. We removed bot/test traffic & very short prompts + used an LLM to filter non-English or non-scientific requests & prompts with personal info, leaving 90K queries.

From selected queries in the 90K pool, we generated cited reports. Quality filtering left 39.5K examples for supervised fine-tuning (SFT). We used separate queries to build ~6K report pairs where two judge models agreed, then trained with direct preference optimization (DPO).

AstaBrief is built for speed & quality. Across Asta’s full pipeline, Fast mode averages 51.1s per report vs. Claude-powered Thinking mode’s 178.5s—about 3.5× faster. And in our evaluations, AstaBrief is competitive with Thinking mode & DR Tulu on answer + citation measures.

Our example repo provides a starting point for generating reports from your own PDFs locally. For more info, read our blog. 💻 Example repo: 📊 Training data: 📝 Blog:

wow, downloading now. You are awesome!

Congratulations! I'll take it

Impressive & useful work ! Bravo to the team ! 👏👏

带引用的报告最怕引用幻觉，8B 在长文献上很容易张冠李戴。关键看它能不能把每句话回指到原文片段，而不是编一个像模像样的 DOI。开源权重的好处是这层能自己审。

Is this a niche-RAG workflow? I supply a bunch of research papers/literature > Ask AstaBrief questions about this topic > It leverages that literature to answer?

base model is Qwen3-8B, a member of the family I run in production. citation precision going from 76 to 90 over the base with sft plus offline dpo is a real jump.

Open weights and open training data is a good combo here. Is the 8B trained to refuse when the excerpts don't support an answer? You'd get citations either way, so I'd want to see how it handles that case.

local + cited reports is a nice combo. i'd stress-test citation entailment more than citation presence though — research agents love attaching a perfectly real paper to a sentence the paper never actually supports

Good job! Can't wait to try it

I am RichardKCollin2, The Internet Foundation. My questions and concerns would not fit in a simple reply so I wrote it into an x article, titled "How People Begin To Ask Matters".