AstaBrief 8B
Ai2's open-weights 8B model (Qwen3-8B-based) that turns a research question plus retrieved literature excerpts into a fully cited scientific report in a single forward pass, about 3.5x faster than its Claude-powered pipeline.
Why it matters
A small open model that matches a proprietary Claude-powered report pipeline (51.1s vs 178.5s per report) with weights and training data released under Apache 2.0; the key lever for citation quality turned out to be a simple citation-density data filter, not a fancier training algorithm.
What you could build with it
An independent researcher or small research team could build a private literature-review assistant on AstaBrief: feed it a lab's own PDF collection and internal notes and get fast, fully cited draft reports that never leave the firewall, since the weights are Apache 2.0 and small enough for a single GPU.
Does it hold up?
Too early to judge for independent deployments; the only production evidence is inside Asta itself, where Ai2 reports Fast-mode positive feedback at a similar rate to the Claude-powered Thinking mode (84.2% vs 85.2%).
Built with AstaBrief 8B
- AstaBrief 8B: How AllenAI Trained a Small Open Model to Generate Cited Scientific Reportsdev.to · Independent breakdown of how Ai2 trained AstaBrief, spotlighting the data-filtering lesson (citation-density filtering beat fancier algorithms).
- AllenAI open-sources AstaBrief — fast reports, open for buildersyoutube · News-style video summary of the AstaBrief release, published within hours of the announcement.
- Ai2 Open-Sources AstaBrief 8B for Fast Scientific Report Generation (Unite.AI)article · Third-party deep-dive on how Ai2 trained AstaBrief 8B, its citation-density filtering lesson, and the 51.1s vs 178.5s Fast vs Thinking mode numbers; notes the example workflow in Ai2's ai2-scholarqa-lib repo for generating reports from own PDFs.
Learn more
First spotted on hf: source.
More AI models
ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69Mistral Forge: enterprise platform for training and continuously improving proprietary modelsMistral AI launched Forge (Sep 28) — a full-lifecycle model training platform (pre-training, SFT, DPO/ODPO, RL…models · JEV 0.69kev: open, trainable Jev-like family of small decision models on Qwen3.5/3.8A family of small decision models (0.8B to 27B) built on Qwen3.5/3.8 that reproduces Jev's prefill-only typed-decision…models · JEV 0.68GPT-Synopsys: OpenAI and Synopsys multi-year deal to build an AI model that operates EDA toolsOpenAI and Synopsys signed a multi-year agreement on Sep 30 to jointly develop GPT-Synopsys, a specialized model that…models · JEV 0.66
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.