← Back to feed

The Allen Institute for AI on October 2 open-sourced AstaBrief 8B, a model built for one narrow task: turning a research question and a set of retrieved literature excerpts into a cited scientific report. The model weights and training data are available on HuggingFace. It is already live in production as Fast mode in Asta, Ai2’s agentic scientific research platform.

What It Does

AstaBrief takes two inputs — a research question and literature passages retrieved from prior search — and outputs a structured, cited report. The key design decision is that it writes the full report in a single forward pass rather than section by section, which previous pipelines did. That change accounts for most of the speed difference.

Across the full Asta pipeline:

• Fast mode (AstaBrief 8B): 51.1 seconds per report
• Thinking mode (Claude-powered): 178.5 seconds per report
• Speedup: 3.5x

Ai2 describes the improvement against proprietary models it was previously using as “nearly an order-of-magnitude reduction in report generation time.”

Why an 8B Model

The 8B parameter scale was a deliberate choice. Ai2 wanted to validate whether a small model trained specifically for scientific report generation could match the quality of general-purpose frontier models, at a fraction of the cost and latency.

Training required tens of thousands of real research queries drawn from Asta user sessions. The data pipeline applied citation-focused filtering and preference data to align the model toward scientific grounding — preserving what the evidence actually supports rather than generalizing beyond it. Hallucinated citations or quietly broadened conclusions are a known failure mode when using general-purpose models on scientific literature.

Local Deployment for Sensitive Research

Open weights matter for research institutions where the query itself reveals unpublished or sensitive work. Running AstaBrief on local infrastructure means the research question never leaves the institution. Ai2 is releasing an example workflow that researchers can adapt to generate reports directly from their own PDFs, using its ai2-scholarqa-lib library.

Asta Integration

Asta now presents two modes in its Generate a report feature. Fast mode runs AstaBrief 8B. Thinking mode runs a Claude-backed pipeline that prioritizes reasoning depth over speed. Users can choose based on whether they need a quick synthesis or a slower, heavier analysis.

The model was trained and evaluated against the frontier as it stood in 2025. Ai2 notes it has not rerun the full evaluation against current frontier models, so quality comparisons against today’s Claude or GPT-6 lineage are not available. The speed and cost advantages remain independent of that comparison.

Context

Ai2 positions Asta alongside platforms like OpenScholar, Elicit, and consensus.ai in the category of AI-assisted scientific literature tools. The distinguishing factor for AstaBrief is the combination of open weights, single-pass generation architecture, and training on actual research queries rather than synthetic tasks. The model is available at huggingface.co/allenai/AstaBrief_8B.