BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
BotPsy is a 0.8 B‑parameter agentic language model built on Qwen3.5‑0.8B‑Think that doubles the base model’s tool‑use capability, achieving state‑of‑the‑art performance for sub‑billion models on six agentic benchmarks. It outperforms every model up to 1.2 B parameters and even beats its larger 2 B sibling on five of six tests, while being designed for on‑device, privacy‑preserving personal‑assistant use. For engineers, BotPsy offers a cheap, self‑hosted alternative to cloud LLMs for low‑to‑medium difficulty tasks that require function calling and multi‑turn tool interaction.
⚡ Key Takeaways
- BFCL v4 score rises from 23.4 (base) to 44.0 with BotPsy.
- ACEBench score rises from 30.1 to 58.7 with BotPsy.
- τ‑bench score rises from 23.2 to 51.5 (mean of retail, telecom, banking subsets).
- BotPsy achieves the highest scores on five of six benchmarks (BFCL v4, ACEBench, τ‑bench, ToolSandbox, BrowseComp‑Plus) in the ≤1.2 B‑parameter class.
- On τ²‑bench telecom scenario BotPsy scores 82.5, surpassing Qwen3.5‑2B‑Think’s 69.3.
Engineers can replace costly cloud APIs with a self‑hosted 0.8 B model that delivers comparable tool‑use performance, reducing operational expense and eliminating data‑privacy risks for on‑device assistants.
✅ Practical Steps
- Download the BotPsy safetensors model from the provided 🤗 Model link and load it with your preferred inference framework.
- Deploy the model on‑device using the GGUF formats (bf16, Q8, Q4) to match your hardware constraints.
- Integrate BotPsy into your agentic pipeline by enabling function‑calling support and routing tool calls through the model’s API.
Want the full story? Read the original article.
Read on Hugging Face Blog ↗