BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
Web page · https://huggingface.co/blog/qvac/botpsy
First seen by Oriel on .
AI summary
BotPsy is a 0.8B-parameter, Qwen3.5-0.8B-based model for on-device tool use that outperforms tested models up to 1.2B parameters across most agentic benchmarks.
- It roughly doubles its base model’s scores on BFCL v4 (44.0 vs 23.4), ACEBench (58.7 vs 30.1), and τ-bench (51.5 vs 23.2).
- BotPsy leads five of six tool-use benchmarks, including several against Qwen3.5-2B-Think, which is 2.5 times larger; Qwen3.5-2B-Think leads on Claw-Eval.
- Training uses supervised fine-tuning, multi-teacher on-policy distillation, and two reinforcement-learning stages with tool-use, instruction-following, and general-domain tasks.
- The model improves instruction following and multi-turn tool use, but performance drops as instructions accumulate and remains weak in lower-resource languages and open-ended search.
- The article says weights will be released under Apache 2.0 in bf16, Q8, and Q4 formats; it cautions that the model can make mistakes and is not intended for high-stakes use.
#language models #on-device ai #llm agents #function calling
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Blog | https://huggingface.co/blog/qvac/botpsy | https://huggingface.co/blog/qvac/botpsy Read discussion |