Sage-A1.5

Sage-A1.5

Works through long scientific tasks as an agent that delegates to subagents, checks its own progress and revises its plan. A 35B mixture-of-experts model for research and engineering workflows.

Made by InternScience on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Sage-A1.5, from InternScience, is an agent model with 35B parameters in a mixture-of-experts design. Its starting points are Agents-A1, from Shanghai AI Laboratory, and Qwen3.6. It targets long scientific tasks. It can split work across subagents, check its own progress and change its approach in the middle of a run.

Who it suits: Researchers and developers who build agents for long, multi-step investigations and want a model that plans, uses tools and double-checks itself.

Industry | Who uses it | How they use it | Example
Science | Researchers | Run exploratory research with results you can inspect | Investigate an open question with several subagents
Software | Agent developers | Split engineering tasks across parallel subagents | Delegate parts of a larger coding job
Professional services | Analysts | Answer professional knowledge questions with tools | Combine a search tool with a code interpreter

If you only need quick chat replies, you can skip it.

Where to use it:

• Hugging Face Transformers
• vLLM and SGLang serving
• Agent frameworks that call functions and tools
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Parallel investigation | Researchers | Subagent findings merged into one result
Self-review | Agent builders | A revised direction after the model reviews its own work
Tool use | Developers | Calls to APIs, code interpreters and search engines

How to set it up:

• Python with pip
• Hardware that holds about 71.9 GB in BF16, or about 18 GB at 4-bit by Hugging Face's estimate
• A serving engine such as Transformers, vLLM or SGLang

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load them in Transformers, or serve them with vLLM or SGLang.

pip install -U huggingface_hub
hf download InternScience/Sage-A1.5

How to use it: Give the model a long task and the tools it may call. Orchestration lets it delegate to subagents with separate contexts. Verification mode is opt-in, and makes the model pause to review what has worked and what has failed.

Tool | What it does
Agentic orchestration | Runs parallel investigations across subagents with their own contexts
Verification mode | Lets the model review its own progress inside its reasoning
Function calling | Connects the model to APIs, code interpreters and search engines

Example: You give it an open research question along with a search tool and a code interpreter. It hands parts of the question to subagents, they test each other's assumptions against evidence, and the model combines their findings into one answer.

For newcomers and experts: If you are new, start with a single focused question and one tool. Turn on verification mode once the basic run works.

The model learns to organize work, not only to solve single tasks. Subagent contexts stay isolated, which reduces interference between concurrent runs. Verification runs inside the chain of thought with no external critic, so it can spot an unproductive search path and switch to another.

A long task and a set of tools arrive
The model splits work across subagents
Each subagent investigates on its own
Verification checks progress and evidence
Findings merge into one result

Good to know:

• Verification mode is opt-in, so it is off unless you turn it on
• The weights work with Transformers, vLLM and SGLang
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: your prompts and any tools or APIs you connect, plus the files downloaded from Hugging Face
• Not for: the card names no banned uses

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 36B
• Weights: about 71.9 GB in BF16; about 18 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 22,224
• Likes: 404
• License: apache-2.0
• Quantized versions: 45
• Fine-tunes: 9
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, vision, agentic, code, tool-use, text-generation, conversational

Links

• Model: https://huggingface.co/InternScience/Sage-A1.5
• Files: https://huggingface.co/InternScience/Sage-A1.5/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download InternScience/Sage-A1.5

Then follow the usage in its model card. Nothing in this kit runs by itself.