Qwen3.5-4B-Router-1M

Qwen3.5-4B-Router-1M

Small decider that returns a calibrated probability for what an agent should do next and which tool it should call. For developers building agents and support bots.

Made by MaziyarPanahi 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: This 4B model decides what an AI agent should do next and which tool it should call. It does not write free text. For every option it gives a calibrated probability. Its base is Qwen3.5-4B with an Unsloth decision head, and it learned from 17,600 labelled rows.

Who it suits: It suits developers building agents or support bots who want a small local model to pick the next step or the right tool.

Industry | Who uses it | How they use it | Example
Software | Agent builders | Pick the next action in an agent loop | Choosing between answering and calling a tool
Customer support | Support platform teams | Route a customer to the right tool or a human | Transferring a request to a human agent
E-commerce | Service desk developers | Choose a lookup tool for a customer question | Calling an order status tool

You can skip it if you need a judge of whole agent traces, free-text replies, or clinical agents.

Where to use it: It runs as Python code on your own machine through Unsloth, with a LoRA adapter on Qwen3.5-4B.

• Python scripts that use the Unsloth decision model class
• Agent loops that need a ranked choice among tools
• Windows, macOS or Linux with a suitable GPU

Use case | Who | What they get back
Next action choice | Agent developers | A probability for each possible next step
Tool choice | Support bot builders | A probability for each available tool
Human handoff | Service teams | A signal to transfer a request to a person

How to set it up: You need Python, the Hugging Face tool and the Unsloth library. The sample code loads the model by its repository name.

• Python
• The Unsloth library
• Disk and GPU memory for a 4B model

1. Install the Hugging Face download tool.
2. Download the adapter files, or let the script fetch them by name.
3. Describe your tools in a question, put the conversation in a state string, then call predict.

pip install -U huggingface_hub
hf download MaziyarPanahi/Qwen3.5-4B-Router-1M

How to use it: Write a question that lists each tool as an option, add the available tools and the conversation to the state, and call predict. You get a probability per option, so you can set your own confidence cut.

from unsloth import FastDecisionModel

model, tokenizer = FastDecisionModel.from_pretrained("MaziyarPanahi/Qwen3.5-4B-Router-1M", max_seq_length=4096, load_in_4bit=False)
FastDecisionModel.for_inference(model)
question = {"decision": {"type": "choice",
    "instructions": "Which available tool should be called next?",
    "criteria": {"get_order_status": "Call `get_order_status`.",
                 "transfer_to_human_agent": "Call `transfer_to_human_agent`."}}}
state = '{"available_tools": [{"name": "get_order_status"}, {"name": "transfer_to_human_agent"}], '\
        '"conversation": [{"role": "user", "content": "I would like to speak to a person please."}]}'
print(FastDecisionModel.predict(model, tokenizer, state, question))

Tool | What it does
FastDecisionModel.from_pretrained | Loads the model and tokenizer
FastDecisionModel.for_inference | Prepares the model for predictions
FastDecisionModel.predict | Returns a probability for each option

Example: A customer writes that they would like to speak to a person. With two tools on offer, an order status lookup and a human transfer, the model returns a probability for each, and the transfer option should come out ahead.

For newcomers and experts: If you are new, think of it as a signpost for an agent that points to the next move. Run the sample and see how it scores two tools against one message.

If you know the field, it is a LoRA of rank 16 trained on four public decision datasets using only their source labels, not the earlier Jev answers. On 480 held-out rows it beat Jev 1.13 by 5.8 points overall. The labels record what the source trace did next, so read the scores as agreement with the source.

Conversation and tool list
Decision head reads the state
Probability for each tool or action
You pick the highest or set a cut

Good to know: The test covered four task families of 120 rows each. Training stopped early on a time budget.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named, but you need a GPU able to run a 4B model
• Data it reaches: the conversation text you pass in on your machine, plus files downloaded from Hugging Face
• Not for: clinical agents or any decision about a patient, judging whole traces, and inputs beyond 4,096 tokens

Title: Overall accuracy on 480 test rows
This model: 86.5%
Jev 1.13: 80.6%

On Hugging Face, 13 November 2026

• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 47
• Likes: 3
• License: apache-2.0
• Based on: unsloth/Qwen3.5-4B
• Tags: peft, safetensors, unsloth, decision-model, tool-routing, agents

Links

• Model: https://huggingface.co/MaziyarPanahi/Qwen3.5-4B-Router-1M
• Files: https://huggingface.co/MaziyarPanahi/Qwen3.5-4B-Router-1M/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download MaziyarPanahi/Qwen3.5-4B-Router-1M

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