The Pulse
Ai2 Releases AstaBrief, an 8B Model for Faster Scientific Reports
Ai2 released AstaBrief 8B, an open-weights model that turns research questions and retrieved papers into cited reports. In Ai2’s tests, its Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for the Claude-powered Think
AI.info Team · Written with AI ·
Ai2’s new scientific-report model takes on a tradeoff researchers already face in its Asta platform: get a cited synthesis sooner with a smaller model, or use the slower Claude-powered mode. Released on October 2, AstaBrief 8B writes a report from a research question and retrieved literature excerpts. Ai2 says its Fast mode averaged 51.1 seconds per report across the full Asta pipeline, versus 178.5 seconds for Thinking mode—a roughly 3.5-fold difference.
The model is now available in Asta’s “Generate a report” feature, and Ai2 has released its weights and training data. The company presents AstaBrief as a faster option, not a replacement for Thinking mode, which it retains for more compute-intensive tasks. The model is based on Qwen3-8B and is released under the Apache 2.0 license.
A single pass replaces a longer pipeline
AstaBrief generates a complete report in one pass from a user’s query and relevant retrieved passages. That design skips the summarization and clustering stages used by Asta’s Claude-based Thinking mode, and avoids building the report section by section. Ai2 says the change speeds report generation without sacrificing performance in the evaluations it ran.
The model’s training combines supervised fine-tuning with direct preference optimization, rather than reinforcement learning. Ai2 says it chose that simpler approach partly because reinforcement-learning methods can be expensive and unstable, and focused instead on the quality of its examples. The release builds on 90,000 research-focused queries filtered from user logs: the team generated reports from them and kept 47,000 examples for supervised fine-tuning, then assembled a separate preference dataset of about 6,000 examples.
Citation support gets its own test
Ai2 evaluated the model on SQABench-CS2, a set of 200 user-written computer-science research questions. Its measures separated report coverage and relevance from citation precision—whether a citation supports the claim attached to it—and citation recall, or whether claims in the report have support in the citations provided. The team also tested AstaBrief on DeepScholarBench, a 63-query benchmark built from recent arXiv papers.
Data filtering shaped the results: Ai2 found that removing training reports with low citation density produced the strongest gains among the filters it tested. The company also warns that a citation can point to a relevant study while the report still overstates what that study established. Its development metrics centered on relevance, coverage and citation grounding; the post says evaluations should also test whether a model preserves the scope and strength of its sources.
Human reviewers favored different strengths
In a small human comparison, three scientific researchers each contributed four or five questions and ranked reports from AstaBrief, the Claude-powered pipeline and Ai2’s DR-Tulu model. DR-Tulu won on overall preference. Two of the three researchers favored AstaBrief on citation-accuracy measures, according to Ai2.
The results do not establish that AstaBrief now matches the best available commercial systems. Ai2 says most of the training and evaluation work took place in 2025, and the team has not rerun the full comparison against current frontier models. “We haven’t rerun the full evaluation against today’s frontier models; the results below are best read as evidence about the particular training and system design choices we tested,” Ai2 writes.
Local deployment, with limits
Open weights let institutions run AstaBrief on their own infrastructure, including behind a firewall, rather than sending report-generation requests to a proprietary model API. Ai2 also released an example workflow for making reports from a researcher’s own PDFs. That option may matter when queries include unpublished or sensitive work; it does not, by itself, make the system’s generated claims reliable without review.
Early use inside Asta offers a limited signal of how the two modes fit together. Among 374 users who tried Fast mode, 29.1% used it on at least two days; 23% continued with Fast mode and did not switch back to Thinking mode for later report threads. Ai2 says user feedback was too sparse for strong conclusions, and reports similar positive-feedback rates for Fast and Thinking—84.2% and 85.2%, respectively.
Researchers can download AstaBrief 8B from Hugging Face or try Fast mode in Asta. The performance comparison remains anchored to Ai2’s older evaluation set and 2025 development work, not a fresh contest against models available in October 2026.