BotPsy-0.8B

BotPsy-0.8B

A small 0.8B model for function calling and agent workflows, built on Qwen3.5-0.8B and light enough to run on a device, for developers who want a local assistant that uses tools.

Made by qvac 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: BotPsy-0.8B is a small model for function calling and agent workflows, built on Qwen3.5-0.8B by Tether AI Research. It acts as a personal assistant that uses tools for simple tasks, and it is light enough for on-device use. This repository holds the BF16 weights in safetensors form, and the GGUF versions sit in a separate repository.

Who it suits: Developers who want a tiny tool-using model on their own hardware will find it fits. The card lists search, travel, commerce, customer operations, device control, files, calendar and messaging as its areas. It suits easy to medium tasks.

Industry | Who uses it | How they use it | Example
Software | Agent builders | Run a local model that chooses which tool to call | A prototype assistant with no hosted API
Customer operations | Support engineers | Handle multi-turn requests that follow a set policy | A telecom support workflow
Devices and vehicles | Embedded developers | Turn spoken or typed requests into device actions | Controlling a connected appliance
Research | ML students | Compare a small agent model with larger ones | Reading the card's benchmark tables

If your tasks are hard, you can skip it, since the card says the model suits low to medium difficulty.

Where to use it:

• Python with the transformers library, which Hugging Face lists
• llama.cpp and the QVAC SDK, through the separate GGUF repository
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Tool calling | Agent builders | A call to the tool that fits the request
Multi-turn tasks | Support engineers | Replies that follow a policy across a long chat
Declining requests | Product teams | A refusal when no suitable tool exists

How to set it up: You need Python with pip and a machine that can hold about 1.7 GB of BF16 weights. The source shows no run code, so the card's own page has the details.

• Python with pip, on Windows, macOS or Linux
• The transformers library
• About 1.7 GB of space for the BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load the model in transformers, with thinking mode on.

pip install -U huggingface_hub
hf download qvac/BotPsy-0.8B

How to use it: Give the model a task together with descriptions of the tools it may call. The card says it thinks before it answers, using a setting named enable_thinking. The source shows no call, so follow the model page for the exact code.

Tool | What it does
enable_thinking=True | Turns on the thinking mode the card describes

Example: You describe a calendar tool and ask the model to book a meeting for Friday. It thinks first, then replies with a call to that tool that carries the details of the request.

For newcomers and experts: Function calling means the model answers with a request to run a tool, rather than only text. A good first step is to try it on one simple, single-tool task before you build anything longer.

The model uses the Qwen3.5 architecture, a hybrid of Gated DeltaNet and gated attention, with text only. On the BFCL v4 test it climbs from the base model's 23.4 to 44.0. A separate check of whether it refuses when no tool fits gives it 86.1. Instruction following improves too, with IFEval going from 59.7 to 90.8. These scores come from the authors.

You give it a task and tool descriptions
The model thinks through the request
It answers or calls the tool that fits
Your code runs the tool and sends the result back

Title: Agent benchmark scores, base model versus BotPsy
BFCL v4 base: 23.4 points
BFCL v4 BotPsy: 44.0 points
ACEBench base: 30.1 points
ACEBench BotPsy: 58.7 points
tau-bench base: 23.2 points
tau-bench BotPsy: 51.5 points

Good to know:

• License: Apache 2.0, as listed on Hugging Face
• Cost to run: none named
• Data it reaches: whatever tools you connect, such as email or a calendar. The source names no direct access
• English is the recommended language
• Not for: hard problems, since the card says it suits low to medium difficulty tasks
• It can still make mistakes, as the card itself notes

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 853M
• Weights: about 1.7 GB in BF16; about 426 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 215
• Likes: 14
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 5
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, tool-calling, function-calling, agent, agentic, on-device, edge, tether-ai, qvac

Links

• Model: https://huggingface.co/qvac/BotPsy-0.8B
• Files: https://huggingface.co/qvac/BotPsy-0.8B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download qvac/BotPsy-0.8B

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