Ai2 open-sources AstaBrief, a fast cited report model

The Allen Institute for AI open-sourced AstaBrief 8B, the fast report generation model inside its Asta platform, along with its training data. The model takes a research question plus retrieved literature excerpts and returns a fully cited report. It is initialized from Qwen3-8B and ships under Apache 2.0, with weights at allenai/AstaBrief_8B.

The speed claim is the headline: 51.1 seconds per report against 178.5 seconds for Asta’s existing Thinking mode, roughly 3.5 times faster, achieved by generating the whole report in one pass rather than section by section.

The training recipe is deliberately plain. Ai2 used SFT plus DPO only, and says it explicitly considered RL, citing DR Tulu, before rejecting it as unstable and expensive. Of 90,000 real Asta research queries surviving quality and privacy filtering, 47,000 became SFT examples, with targets generated by a pipeline mixing Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1. A disjoint query subset produced about 6,000 DPO pairs, judged by GPT-4.1 and DeepSeek-R1 with only pairs where both agreed kept, at 95% agreement with human preference.

The negative result is the useful part

Ai2 tested four statistical filters on the training data: output-to-input token ratio, citation relevance, citation density and citation diversity. Filtering on low citation density alone produced by far the biggest quality gain. Combinations of filters and learning rate sweeps added nothing meaningful.

On evaluation, SQABench-CS2 covers 200 user-written computer science questions, and DeepScholarBench scores 53.50 for AstaBrief against 60.25 for the full Asta ScholarQA pipeline and 56.26 for DR-Tulu-8B. LLM-judged win rate against the Claude pipeline was 55% on dev and 72% on test.

Ai2 is direct about the limits. Most of the training and evaluation was completed in 2025 and has not been rerun against current frontier models, so the comparisons describe a 2025 baseline rather than October 2026. The human study was 14 questions with 3 researchers, DR-Tulu won overall preference there, and the usage telemetry comes from 374 users, which Ai2 calls too sparse to draw strong conclusions from. The metrics also test whether a citation exists, not whether a model preserves the evidentiary scope of its sources.

(Original source: Allen Institute for AI)