On 21 October 2026, ModelScope’s account on X announced Sage-A1.5, an open model for “long-horizon scientific workflows”. According to its model card, the model comes from Shanghai AI Laboratory, which posted a technical report on it on 20 October 2026. It is published under the name InternScience on Hugging Face and on ModelScope, the model-sharing site that announced it, and is released under the Apache 2.0 open-source license. This article explains what the model is for, how it compares and how to try it.
What it is built for
Long-horizon research runs over many steps: searching published papers, running an analysis, following a lead, hitting a dead end, and finally pulling the evidence together. Sage-A1.5 is built for that kind of work. ModelScope’s announcement highlights three features:
- A large working memory. The model can handle about 262,000 tokens at once (the 256K in ModelScope’s announcement). A token is a short chunk of text, often part of a word.
- Splitting the work. In a mode called agentic orchestration, the model hands separate lines of inquiry to helper agents that work at the same time, each with its own separate memory, then combines what they find. The model card says this is off by default: developers switch it on, and their own software runs each helper.
- Checking itself. In an optional Verification Mode, the model reviews its progress, checks its evidence and redirects searches that have stalled. ModelScope says this happens “within the model’s own reasoning, without a separate critic model at inference”. In other words, no second model checks its work while it runs.
ModelScope says the model can also use outside software through APIs (connections between programs), code interpreters (programs that run code) and search engines, and the model card lists it as a vision-language model, so it reads images as well as text. ModelScope says it aims to go beyond “predefined question answering” and produce “inspectable research artifacts”.
Is there anything similar?
Yes. Several AI services and open projects also take on long research tasks, though they work differently.
- Deep-research tools in chat apps. OpenAI’s deep research, launched on 2 February 2025, is an agent in ChatGPT that searches the web, analyses what it finds and writes a cited report. Google added Deep Research to Gemini on 11 December 2024: it drafts a research plan for the user to approve, browses, and produces a report that can be exported to Google Docs. Perplexity’s Deep Research, launched on 14 February 2025, runs dozens of searches and writes a report, finishing most tasks in under three minutes, Perplexity says. All three are finished services that run on their makers’ servers.
- Open research agents. Hugging Face published an open-source version of a deep-research agent in February 2025, and LangChain released Open Deep Research, a configurable open-source agent that works with models from many providers; LangChain archived it, making it read-only, on 21 August 2026. These are the software around a model, not models themselves. Sage-A1.5 is the reverse: a model, with the surrounding software left to the user.
- The larger models in its comparison. The makers compare Sage-A1.5 with three models the model card calls “much larger frontier models”. Kimi K3, from Moonshot AI, is a 2.8-trillion-parameter model with open weights, released on 16 July 2026 according to Epoch AI. GLM-5.3, from Z.ai, was released on 14 August 2026 and is aimed mainly at coding and long tasks. GPT-5.6 Sol was OpenAI’s flagship model when it launched on 9 July 2026, and is available through OpenAI’s own products and API rather than as a download; OpenAI now calls GPT-6 Astra its “most powerful model”.
How it compares
The model card gives the full comparison across 22 tests. Sage-A1.5 is ahead of the larger models on 8 of them and behind at least one of them on the other 14. The table shows the three wins ModelScope highlighted and four of the losses:
| Test | Sage-A1.5 | Highest of the three larger models |
|---|---|---|
| PinchBench | 90.0 | 88.2 |
| FrontierScience-Olympiad | 81.0 | 73.0 |
| FrontierScience-Research | 48.3 | 26.7 |
| SWE-bench Pro (software engineering) | 61.2 | 64.6 |
| HLE with tools | 55.1 | 64.5 |
| CritPt | 20.0 | 32.3 |
| ChemCOTBench | 66.5 | 82.5 |
ModelScope calls the results “SOTA across scientific benchmarks”; SOTA is short for state of the art, the best result reported so far. On the makers’ own figures, the clearest wins are on research-style science tests, above all FrontierScience-Research, where Sage-A1.5 scores 48.3 against 26.7 for the best larger model. PinchBench is also a win, but the model card groups it with general agent tasks rather than science tasks. The losses include all four software-engineering tests, among them SWE-bench Pro; two of the four long-horizon search tests, BrowseComp and DeepResearch-Bench2; hard expert questions answered with tools (HLE with tools); and the specialist science tests CritPt and ChemCOTBench. On SWE-bench Pro, Sage-A1.5’s 61.2 is up from 43.6 for the lab’s earlier model, Agents-A1.
Three cautions apply. All these scores come from the makers, who chose which models to compare. The model card says that where a rival had not published a score, the makers ran the test themselves. And neither the announcement nor the model card cites results from other groups.
The trade-offs matter as much as the scores:
- Open and local. Unlike the deep-research services above, the model can be downloaded and run on a user’s own machines under Apache 2.0, so the material it works on does not have to be sent to an outside AI service.
- Smaller, but not tiny. The technical report describes it as an “MoE with 35B total and 3B active parameters”. Parameters are the numbers a model learns in training. MoE, short for mixture of experts, means only about 3 billion of the 35 billion parameters work on each piece of text. That keeps the computing work for each step small, but all 35 billion still have to be loaded into memory, so the computer running it needs room for the whole model.
- Assembly required. The helper-agent mode is off by default, and the user’s own software has to run each helper. Out of the box, it is a model, not a ready-made research app.
Who it is for
Sage-A1.5 is aimed at developers and research groups that want to build their own research assistant: university and company labs that cannot send data to an outside service, teams that want to inspect or adapt the model under an open license, and builders of open research agents who need a capable model to sit inside them. People who simply want a research report today will find the chat-app services above easier, because they need no setup.
How to use it
These steps follow the model card and ModelScope’s announcement:
- Download it. The main model is on Hugging Face and ModelScope as InternScience/Sage-A1.5, in BF16. ModelScope’s announcement also lists smaller formats, which are separate downloads on Hugging Face: an FP8 version (InternScience/Sage-A1.5-FP8) and, for tools that use the GGUF file format, GGUF versions at F16, Q8 and Q4 (InternScience/Sage-A1.5-F16-GGUF, InternScience/Sage-A1.5-Q8_0-GGUF and InternScience/Sage-A1.5-Q4_K_M-GGUF). Q8 and Q4 are compressed and smaller.
- Choose serving software. The model card says the files work with Hugging Face Transformers, vLLM and SGLang, and gives ready-made commands for vLLM and SGLang that start a local server with an OpenAI-style interface, using the full 262,144-token context. vLLM also has a text-only option that skips the image part to save memory.
- Turn on the research modes if needed. Verification Mode is switched on per request with a setting in the chat template. Helper agents are switched on by giving the model a tool called spawn_subagent; the user’s own software then runs each helper and returns its results.
- Check the results. The model card says its scores come from evaluation code the lab has made public, so other groups can rerun tests under the same rules. That code was published with the lab’s earlier model, Agents-A1, and the lab describes it as covering “selected domains”.
Our take
Sage-A1.5 is a serious attempt at a real problem: keeping an AI on track across a long investigation. Its strongest selling point is not any single score but the combination of open weights, a permissive license and research-specific features in a model that is far smaller than the rivals whose sizes are published. For a lab that wants to keep its data in-house, that is a genuinely useful option.
We would not read the table as proof that it beats frontier models at research. The headline wins are on tests the makers chose, some rival scores were produced by the makers themselves, and the model trails on coding and expert-knowledge tests that matter for real lab work. Its most distinctive feature, the helper agents, also needs engineering work before it does anything. Our advice: worth a trial for technical teams building their own research tools, ideally on a real project from their own field; not yet a replacement for the deep-research services for everyone else. Results from other groups will be the real test.
Correction, 21 October 2026: an earlier version of this article gave Kimi K3’s release date as 14 July 2026 (it was 16 July), said LangChain maintains Open Deep Research (LangChain archived it on 21 August 2026), said the model card shows Sage-A1.5 ahead on three tests (it is ahead on 8 of 22) and gave its working memory as about 256,000 tokens (it is 262,144). It also named a base model for Sage-A1.5, which the makers have since removed from the model card.
Correction, 22 October 2026: an earlier version of this article said GPT-5.6 Sol is OpenAI’s flagship model. It now says GPT-5.6 Sol was OpenAI’s flagship model when it launched on 9 July 2026, and that OpenAI now calls GPT-6 Astra its most powerful model.