# Sage-A1.5 ![Organization](https://img.shields.io/badge/Org-Shanghai%20AI%20Lab-555) ![Context](https://img.shields.io/badge/Context-262144%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **Quick answer:** Sage-A1.5 is a 35B-parameter Mixture-of-Experts agentic model from Shanghai AI Laboratory (InternScience), built for long-horizon scientific problem solving. It scores 95.1 on GAIA — higher than much larger frontier systems it's compared against (Kimi-K3, GLM-5.3, GPT-5.6-sol) — and leads comparable ~35B-class models on BrowseComp, SWE-bench Pro, SWE-bench Multilingual, PinchBench, and several science-reasoning benchmarks. Apache 2.0 open weights. ## At a Glance **Where Sage-A1.5 leads** - 95.1 on GAIA — highest of every model compared, including larger frontier systems - First among ~35B-class peers on BrowseComp (81.7–82.8), SWE-bench Pro (61.2), MLE-Lite (65.2), AA-LCR (81.0), and PinchBench (90.0) - Opt-in "Agentic Orchestration" mode fans out sub-questions to parallel sub-agents with isolated contexts; opt-in "Verification Mode" lets the model self-review mid-trajectory **Where it lags** - Trails larger frontier models (Kimi-K3, GLM-5.3, GPT-5.6-sol) on several engineering and tool-heavy benchmarks (HLE w/ tools, VitaBench, ChemCOTBench) - 35B active weight class requires meaningful GPU infrastructure to self-host at full 262K context **Best for:** teams running long-horizon agentic research/search workflows who want an open-weight model competitive with frontier-scale systems at a fraction of the parameter count. ## What Sage-A1.5 Is Sage-A1.5 is a 35B Mixture-of-Experts model from Shanghai AI Laboratory's InternScience group, built on the Qwen3.5-MoE architecture with native vision-language support. It extends the earlier Agents-A1 model with three new capabilities: agentic orchestration (parallel sub-agent task delegation with isolated contexts), agentic verification (self-review and strategy revision mid-trajectory), and open scientific discovery (exploratory research beyond fixed Q&A). The model is positioned specifically for long-horizon, tool-using scientific and engineering workflows — search, coding agents, document QA — rather than as a general chat assistant, and its released benchmark suite reflects that: agentic search (GAIA, BrowseComp), engineering (SWE-bench Pro/Multilingual, MLE-Lite), and PhD-level science reasoning (FrontierScience-Olympiad/Research, CritPt). ## Specifications | Field | Value | |---|---| | Organization | Shanghai AI Laboratory (InternScience) | | Total / active parameters | ~35B (MoE; qwen3_5_moe architecture) | | Context length | 262,144 tokens | | Modality | Text (+ vision-language variant) | | License | Apache 2.0 | | Release date | October 2026 | ## State-of-the-Art Results | Benchmark | Sage-A1.5 | Notes | |---|---|---| | GAIA | 95.1 | Highest of all compared models (incl. larger frontier systems) | | BrowseComp | 81.7 (w/o orchestration) / 82.8 (w/ orchestration) | Best among ~35B peers | | SWE-bench Pro | 61.2 | Best among ~35B peers | | SWE-bench Multilingual | 72.7 | Best among ~35B peers | | AA-LCR | 81.0 | Best among ~35B peers | | PinchBench | 90.0 | Best among ~35B peers | | FrontierScience-Research | 48.3 | Best among all compared models | | CritPt | 20.0 | Best among ~35B peers | | OmniDocBench | 91.5 | Best among all compared models | | IFBench | 81.6 | Best among all compared models | *Scores are from InternScience's own technical report (arXiv:2610.28578, Oct 2026), evaluated with the mini-SWE-agent harness for SWE-bench series (temp 1.0, top-p 0.95, 200K context, offline). Benchgen has not independently re-run these evaluations.* **Excluded from this benchmark table (not mapped to a Benchgen entity):** DeepResearch-Bench2, XBench-DS-2510, MLE-Lite, FrontierScience-Olympiad, HiPhO, HLE w/ tools (distinct task setup from Benchgen's generic Humanity's Last Exam entity — not a direct comparison), CMT, Agents' Last Exam (ALE-CLI), VitaBench, ChemCOTBench, MatTools — no existing cross-model-comparable Benchgen entity with confirmed methodology/source found during this scan; flagged for a future review rather than guessed. ## Sage-A1.5 on Benchgen > No Benchgen results yet — [be the first to run Sage-A1.5](/models/internscience/sage-a1-5). ## Frequently Asked Questions

Who built Sage-A1.5?

Shanghai AI Laboratory's InternScience group, as a successor to their earlier Agents-A1 model.

Is Sage-A1.5 open source?

Yes — released under Apache 2.0 with open weights on Hugging Face.

What is Sage-A1.5 best at?

Long-horizon agentic search and research tasks: it scores 95.1 on GAIA, ahead of every other model InternScience compared it against, including much larger frontier systems.

--- *Specs from the Sage-A1.5 Hugging Face model card and technical report (arXiv:2610.28578, InternScience, Oct 2026). Last updated 2026-10-21.*

Sage-A1.5 — Evaluations

Evaluations