Ai2 Open-Sources AstaBrief 8B, a Fast Scientific Report Model

The Allen Institute for AI (Ai2) has open-sourced AstaBrief 8B, a specialized language model that converts a research question and a set of literature excerpts into a fully cited scientific report. The weights, training data and example workflows are all available on Hugging Face under an Apache 2.0 license — meaning anyone can download, inspect and adapt the model at no cost.

AstaBrief is live now as “Fast mode” inside Asta, Ai2’s agentic research platform, sitting alongside a Claude-powered “Thinking mode” for users who want deeper synthesis and don’t mind waiting longer. The speed difference is substantial: Fast mode averages 51.1 seconds per report, compared with 178.5 seconds for Thinking mode — roughly 3.5 times faster.

What Makes It Different From Other AI Research Tools

Scientific AI tools have multiplied quickly, but the market is fragmented. Semantic Scholar handles paper discovery; Elicit and Consensus focus on structured data extraction and evidence synthesis. None of them offer downloadable model weights or local deployment — they are cloud-based, API-dependent services. Elicit’s Pro tier runs $49 per user per month; Consensus Pro costs around $20 per month. On the proprietary side, OpenAI’s Deep Research and Google Gemini Deep Research produce multi-source synthesis reports but carry paid plans starting at $20 per month and cannot be self-hosted.

AstaBrief’s Apache 2.0 license and self-hostability are a genuine differentiator. Because the model is small enough to run on a single consumer GPU, institutions can deploy it entirely behind their own firewall — important when research questions involve sensitive or unpublished data that cannot be sent to a third-party API.

How Ai2 Built It

AstaBrief is built on top of Qwen3-8B through post-training, specifically supervised fine-tuning (SFT) followed by direct preference optimization (DPO). Ai2 explicitly considered reinforcement learning and rejected it in favor of this simpler, more stable pipeline — betting that high-quality training data could do the work that a more complex optimization method might otherwise compensate for.

That bet shaped how the training data was collected. Rather than using synthetic prompts or benchmark-style tasks, Ai2 drew on real queries submitted through Asta by actual scientists. After filtering for quality, relevance and privacy — stripping out bot traffic, short queries, non-English requests and anything containing personal information — the team was left with 90,000 research-focused queries. From those, it generated 47,000 usable SFT training examples using a mix of proprietary models including Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1.

“Rather than training only on synthetic prompts or benchmark-style tasks, we wanted AstaBrief to learn from real queries from real scientists,” wrote Ai2 comms lead Kyle Wiggers.

The DPO stage required preference pairs — two reports per query, with one ranked above the other. Ai2 used approximately 6,000 such pairs, requiring both GPT-4.1 and DeepSeek-R1 to agree on a winner before a pair was included. The team reports 95% agreement between those LLM judges and human preferences.

The model also generates reports in a single forward pass rather than writing section by section, which is a key reason for the speed gain. The Claude-powered Thinking mode runs snippet summarization and clustering before writing; AstaBrief skips those intermediate stages entirely. Ai2 says it found it possible to bypass them without sacrificing report quality — a finding that informed the broader architecture decision.

Evaluation was conducted on SQABench-CS2, a 200-question computer science research benchmark, tracking rubric score, answer precision, citation precision and citation recall. Secondary evaluations used DeepScholarBench (63 questions drawn from recent ArXiv papers) and two pairwise comparisons against Claude-powered output, including a human study. Ai2 is candid that most of this evaluation was completed in 2025 and has not been rerun against today’s frontier models, so head-to-head comparisons with current proprietary offerings should be read carefully.

What This Means for Students

For students, AstaBrief offers two concrete benefits that don’t require any technical setup to access. First, Fast mode in Asta is free to use today — no subscription required — and produces cited literature reports grounded in retrieved evidence, which is directly useful for literature reviews and research papers. Second, for students in machine learning or natural language processing programs, the released training pipeline is an unusually complete educational resource: 47,000 SFT examples, roughly 6,000 DPO preference pairs and a documented evaluation methodology, all on HuggingFace. Reproducing a project like this from scratch would typically require institutional compute budgets and access to proprietary data that most students don’t have.

The self-hostability angle also matters beyond elite research universities. A lab or institution that cannot afford $49-per-seat subscriptions to cloud-based research tools can run AstaBrief locally for the cost of the hardware — and keep sensitive pre-publication work off third-party servers in the process.

The caveat worth keeping in mind: AstaBrief is an 8-billion-parameter model trained for a specific task. It will not replace the depth of a Claude-powered synthesis for complex, multi-step research questions, and its evaluation benchmarks predate the current generation of frontier models. But as a fast, free, locally deployable starting point for scientific report generation, it fills a gap that nothing else in the open-weights ecosystem currently occupies.

Source: Ai2 / Hugging Face