Intel / article

Open-sourcing AstaBrief, the fast report-generation model in Asta

Source: allenai.org
Date: Published Oct 2, 2026

From the Daily Brief

Ai2 has open-sourced AstaBrief 8B, the model behind the new Fast mode in its Asta research assistant. Given a research question and retrieved literature excerpts, it writes a report with citations. Ai2 says the goal is to approach proprietary report quality while generating faster and serving more cheaply. In Asta, reports take about 51 seconds, compared with roughly 178 seconds for the Claude-powered mode. The recipe is published alongside the weights. Ai2 started from Qwen3-8B, filtered about 90,000 real user queries, ran supervised fine-tuning on reports filtered for citation quality, added DPO preference pairs, and designed the model to write the whole report in a single pass. The training data is open too, and the Hugging Face post includes a local example workflow. For teams building research or retrieval tools, this is a reproducible template for distilling a narrow, citation-heavy task into a small model they can run on their own hardware.

Read the 2026-10-03 Brief →

Context

Ai2 says AstaBrief is faster because it was trained to write a whole cited report in one pass from a question and retrieved excerpts, skipping the snippet summarization and clustering stages that its Claude-powered Thinking mode uses and the section-by-section writing. Across the full Asta pipeline, Ai2 reports average times of 51.1 seconds for Fast mode (AstaBrief) and 178.5 seconds for Thinking mode, about 3.5 times faster. The comparison covers a model and a pipeline that both changed, so it does not isolate either one. Most training and was done in 2025, so Ai2 cautions that the comparisons do not show how AstaBrief ranks against today's frontier models. Development account

Key takeaways · AI-distilled

Ai2 fine-tuned Qwen3-8B with SFT then DPO rather than RL, which it judged unstable and expensive, and trained AstaBrief to write the whole cited report in one pass, skipping the snippet summarization and clustering stages its Claude-based Thinking mode uses.
Training started from about 90K filtered real Asta research queries. SFT used 47K reports generated by proprietary models; the roughly 6K DPO pairs were kept only when two judges, GPT-4.1 and DeepSeek-R1, agreed on the better report.
Of four statistical data filters Ai2 tested, dropping synthetic reports with low citation density gave the strongest gains; more aggressive filtering, filter combinations and learning-rate sweeps added nothing meaningful, per the post.
In Asta, Fast mode averages 51.1 seconds per report versus 178.5 seconds for Thinking mode, about 3.5x faster. In a small 14-question human study DR Tulu won overall preference, while two of three researchers preferred AstaBrief on citation accuracy.
Ai2 cautions that most training and evaluation finished in 2025 and was not rerun against current frontier models, and that its development metrics centered on relevance, coverage and citation , not on whether a report overstates the scope of its sources.

Terms in this piece · Glossary

eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.

Read the source allenai.org