SSyncDev

AstaBrief 8B

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

AstaBrief 8B is Ai2's open 8B model that writes a cited scientific report in one pass from a research question and retrieved literature excerpts.

Key specifications

Specification score
4.7
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Weights only; also powers Fast mode in Ai2's Asta product.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

Open weights

What is AstaBrief 8B?

Literature review agents are usually slow because they write in pieces: outline, section, section, revise. AstaBrief 8B, from Ai2, skips that. Given a research question and a pile of retrieved excerpts, it writes the entire cited report in a single pass. Ai2 open-sourced it on 2 October 2026, and it already powers Fast mode in the Asta research assistant.

The speed difference is the headline. In Asta's full pipeline, Fast mode averages about 51 seconds per report. Thinking mode, which runs on a much larger Claude-based pipeline, takes about 179 seconds. That is roughly three and a half times faster.

Built small, trained carefully

The base is Qwen3-8B. Most of the work went into data:

  1. Supervised fine-tuning on about 47K examples, filtered from roughly 90K research queries.
  2. Offline DPO on about 6K preference pairs. Two judges, GPT-4.1 and DeepSeek-R1, had to agree on each pair, otherwise it was dropped.
  3. No reinforcement learning.

One lesson from training stands out: filtering SFT examples by citation density gave the biggest gains, while fancier filters and learning-rate sweeps added little. An SFT-only checkpoint (AstaBrief_8B_SFT) and the training data are released too, which makes ablations straightforward.

How good is it?

On SQABench-CS2, 200 computer-science questions scored on rubric coverage, answer precision and citation precision and recall, it was competitive with the Claude-powered pipeline and with DR Tulu. In a small human study (14 questions, three researchers) DR Tulu won on overall preference, yet two of the three researchers preferred AstaBrief on citation accuracy. Among 374 Asta users who tried Fast mode, 84.2% of feedback was positive, against 85.2% for Thinking.

Limits Ai2 is upfront about

  • Training and evaluation finished mostly in 2025 and were not rerun against current frontier models.
  • The main benchmark covers computer science only.
  • Metrics measure relevance and citation grounding, not whether the report overstates what a source shows.
  • It was designed to sit inside Asta's retrieval framework, not to work as a standalone chatbot.

Running it

Field Value
Repository allenai/AstaBrief_8B
Size 8B
Input Research question plus retrieved excerpts (text)
Output Cited report (text)
Licence Apache 2.0 (stated with the weights)
Hosted API None; self-host

Ai2's ScholarQA library includes an example workflow for generating reports from your own PDFs. The repository was created in February 2026, months before the public release; we use the October announcement date.

Trying it on your field

Retrieval quality will decide most of the outcome, so keep your retriever fixed and vary only the writer. Pick twenty questions from your own domain, ideally outside computer science, and have a subject expert check every citation against its excerpt. Compare against a larger hosted research model such as o3-deep-research for quality, and against other summarisation models for cost. More Ai2 releases are on the Ai2 provider page.

How to evaluate AstaBrief 8B

Compare the model on the work you need to complete.

Tasks and inputs

AstaBrief 8B is listed for Question answering, Summarization, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

It writes a complete, cited scientific report in one pass from a research question and retrieved literature excerpts.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

cahya

cahya/t5-base-indonesian-summarization-cased

cahya/t5-base-indonesian-summarization-cased is a source-linked Hugging Face repository assigned to Summarization. Its recorded task interface accepts text and produces text.

— context

cnicu

cnicu/t5-small-booksum

cnicu/t5-small-booksum is a source-linked Hugging Face repository assigned to Summarization. Its recorded task interface accepts text and produces text.

— context

csebuetnlp

csebuetnlp/mT5_multilingual_XLSum

csebuetnlp/mT5_multilingual_XLSum is a source-linked Hugging Face repository assigned to Summarization. Its recorded task interface accepts text and produces text.

— context

d0rj

d0rj/rut5-base-summ

d0rj/rut5-base-summ is a source-linked Hugging Face repository assigned to Summarization. Its recorded task interface accepts text and produces text.

— context