Ai2 open-sources AstaBrief 8B, a Qwen3-8B-based report-generation model, and its training data; now live as Fast mode in Asta

Ai2 has released the weights and training data for AstaBrief 8B on Hugging Face, along with an example workflow for producing cited reports from a user's own PDFs. The model is Qwen3-8B fine-tuned on 47K examples drawn from 90K filtered real queries, then refined with preference training (DPO). It is also live in Asta's Generate a report feature as "Fast mode", next to the Claude-powered "Thinking mode". Ai2's own figures show 51.1 seconds per report against 178.5 seconds for Thinking mode, about 3.5x faster.

Research teams and institutions with sensitive or unpublished material can now run a literature-report model on their own hardware, and they can inspect the training data. That is a real, if narrow, benefit. Developers get a reusable recipe for fine-tuning a small model for cited report writing. The wider impact is limited: this is one task, one small model, and the open release does not shift power away from the large proprietary labs.

The post describes the speed-up as "nearly an order of magnitude", but Ai2's own Asta numbers show about 3.5x. The claim that it matches proprietary quality rests on comparisons run in 2025 against older models, so it says little about current systems. The post also gives no independent evaluation, and the 51.1s average is Ai2's own measurement.

1. 01Independent evaluations of AstaBrief 8B against current proprietary models on citation accuracy and report quality, not Ai2's 2025 comparisons.
2. 02Hugging Face download counts, community fine-tunes, and reports of institutions running it on their own infrastructure for private documents, over the next one to three months.
3. 03Whether Ai2 publishes updated benchmarks or moves more Asta users to Fast mode, and whether usage data or user complaints show quality holding up against Thinking mode.