Agents-A1

Agents-A1

Runs multi-step agent work: planning, search, tool calling and scientific reasoning, as a 35B open-weight language model. Built for developers who wire a model into agents that call tools.

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: Agents-A1 is a language model with 35B parameters and a Mixture of Experts design, tuned for agent work. It targets search over long task chains, software engineering, science, following detailed instructions and calling tools. The repository holds the weights and configuration in Hugging Face Transformers format.

Who it suits: It suits developers and research teams who build agents that plan, search and call tools, and who can host a large model themselves.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Put the model behind an agent loop that calls tools and APIs | A coding helper that reads files and runs a code interpreter
Science | Research groups | Use it for tool-assisted scientific reasoning and professional question answering | A literature search agent that gathers and compares sources
Enterprise IT | Platform teams | Serve it through an inference engine for internal assistants | An internal assistant that follows multi-rule instructions
Education | Lab instructors | Study how an agent plans steps and adapts to results | A class demo that shows a model revising its plan

If you want a small model for a laptop, you can skip this one. The card says a 4B model has been released.

Where to use it:

• Hugging Face Transformers
• vLLM and SGLang, which the card names as compatible
• Windows, macOS or Linux, where you can download the files with Python
• A Mac, through the quantized copies the card credits to the mlx-community

Use case | Who | What they get back
Long search tasks | Research agents | A model that breaks a question into steps and adjusts as results arrive
Tool calling | Developers | Function calls to APIs, code interpreters and search engines
Instruction following | Product teams | Answers that respect several constraints at once

How to set it up: You need Python with pip and a machine with room for a large model.

• Python and pip
• About 70.2 GB of disk for the BF16 weights, or about 17.6 GB at 4-bit, which is an estimate
• A runtime the card names: Transformers, vLLM or SGLang

1. Install the Hugging Face client with pip.
2. Download the repository with the hf command.
3. Load the files in your runtime. The part of the card shown here gives no run commands.

pip install -U huggingface_hub
hf download InternScience/Agents-A1

How to use it: Once the model runs in your engine, you give it a task and connect the tools it should call. The card describes it as splitting a task into smaller steps, planning ahead and changing course when a step returns something unexpected.

Example: You connect a search tool and ask, "Where do sources disagree about this topic?" It plans the searches and calls the tool more than once. Then it reads what comes back and writes up the comparison.

For newcomers and experts: This is a large model tuned to act as an agent, not only to chat. The first small thing to try is the download command above, or one of the quantized copies, which are much smaller.

The card describes three training stages. First comes supervised fine-tuning across all domains. Then one teacher model is trained per domain. Last, one student learns from several teachers at once through on-policy distillation. It also builds training data from actions, observations and checker outcomes. The trade-off is size, since the full weights are large.

Supervised fine-tuning on all domains
One teacher model per domain
Multi-teacher on-policy distillation
Agents-A1 for tool-using agents

Good to know:

• Other versions: Hugging Face lists 76 quantized and 5 fine-tuned versions made by other people.
• Size: the weights are about 35.1B parameters in Safetensors format.

Title: Weight file size
BF16 weights: 70.2 GB
4-bit weights (estimate): 17.6 GB

• License: Apache 2.0, as the License line says.
• Cost to run: no API key, account or paid service is named. You pay for hardware able to hold the weights.
• Data it reaches: whatever tools you connect, such as APIs, code interpreters and search engines. The model alone only reads your prompts.
• Not for: the part of the card shown here lists no forbidden uses.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 35.1B
• Weights: about 70.2 GB in BF16; about 17.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 7,232
• Likes: 669
• License: apache-2.0
• Quantized versions: 76
• Fine-tunes: 5
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, vision, agentic, text-generation, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download InternScience/Agents-A1

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