Ai2 opens AstaBrief 8B, the model that writes Asta's fast cited reports

Ai2 open-sourced AstaBrief 8B on October 2, 2026. The model takes a research question plus retrieved literature excerpts and writes a cited report. AstaBrief already runs as Fast mode in the Generate a report feature of Asta, Ai2's platform for scientific work, next to a Claude-powered Thinking mode. Ai2 published the weights, an earlier checkpoint and the training data on Hugging Face.

What Ai2 released

The AstaBrief collection (opens in a new tab) on Hugging Face has five items: the final AstaBrief_8B model, the AstaBrief_8B_SFT checkpoint, the AstaBrief_SFT_Mix and AstaBrief_DPO_Mix datasets, and AstaBrief_prompts. The AstaBrief_8B model card (opens in a new tab) lists the license as Apache 2.0 and the base model as Qwen3-8B. The card says the model is intended for research and educational use under Ai2's Responsible Use Guidelines.

Ai2 also points to example code for producing reports from your own PDFs. It lives in the scholarqa/lite folder (opens in a new tab) of the ai2-scholarqa-lib repository, where a ScholarQALite class generates the report in one shot instead of running quote extraction and clustering first.

How AstaBrief was trained

Ai2 chose supervised fine-tuning (SFT) followed by direct preference optimization (DPO) and skipped reinforcement learning. The announcement (opens in a new tab) gives the reason: RL can be unstable and expensive, and the team wanted a setup that was easier to debug.

Training started from real Asta user queries. After filtering out bot traffic, very short queries, non-scientific requests and prompts with personal information, 90K research queries remained. Ai2 ran those through its ScholarQA pipeline, backed by Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1, to produce target reports. Quality filtering left 47K usable SFT examples.

For DPO, Ai2 built report pairs from a separate set of queries. GPT-4.1 and DeepSeek-R1 each picked a winner, and only pairs where both judges agreed were kept. Ai2 says the judges matched human preferences 95% of the time. The final DPO set is about 6K examples.

The most useful filter was simple. Removing synthetic reports with low citation density gave the largest gain, while stacking more filters or sweeping learning rates added little. Per the model card, DPO training ran in open-instruct on 8xH100s for 7 epochs.

Speed and quality, as Ai2 reports them

AstaBrief writes the whole report in one pass and skips the snippet summarization and clustering steps that Thinking mode uses. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report against 178.5 seconds for Thinking mode, about 3.5x faster. The post also describes nearly an order-of-magnitude cut in generation time against the proprietary models it tracked; the 3.5x figure is the end-to-end one users see.

On the ScholarQA-CS2 test set, the model card gives AstaBrief-8B an average of 87, against 83.7 for the SFT checkpoint and 77.3 for plain Qwen3-8B. Citation precision rose from 76.2 for Qwen3-8B to 90.5. Against other report systems the picture is mixed. AstaBrief scores 87.0 on the test split against 86.2 for the Claude-backed Asta ScholarQA pipeline and 88.8 for DR-Tulu-8B. On DeepScholarBench it trails, with 53.50 against 60.25 for Asta ScholarQA.

A 14-question human study with three researchers also gave mixed results. DR Tulu won on overall preference, while two of the three researchers preferred AstaBrief on citation accuracy.

Who AstaBrief is for

The clearest case is a lab or institution that cannot send queries to an outside API, for example when a question reveals unpublished work. Ai2 makes this argument itself: open weights let teams run report generation on their own hardware, behind their own firewall.

AstaBrief is also a reference for anyone building retrieval-augmented generation over papers. The model writes; it does not search. Your pipeline has to supply the excerpts, which is the part covered in building a RAG pipeline with Postgres and pgvector.

How to try AstaBrief

The quickest route is Fast mode in Asta's Generate a report feature. To run the model yourself, the card shows a snippet that loads allenai/AstaBrief_8B and generates with vLLM, using temperature 0.7, top_p 0.95 and max_tokens 4096. For a hosted endpoint, see serving an open model with vLLM.

The card warns that the model was fine-tuned on one specific prompt, published as sft_prompt.txt in the AstaBrief_prompts dataset. A different prompt format may give degraded or inconsistent output.

Limits and open questions

Most training and evaluation finished in 2025. Ai2 says it has not rerun the full evaluation against current frontier models, so the comparisons reflect the models available then.

The main benchmark covers computer science questions only, and the model card lists English as the only language. Results for other scientific fields are not reported.

Ai2 is open about what its metrics miss. They check relevance, coverage and whether citations support claims. They do not test whether a report widens a finding beyond its sample or turns a description into a recommendation. Early production numbers are small: of 374 users who tried Fast mode, 29.1% used it on two or more days.