Hugging FaceMilan Gritta, Jihye Back, JinUk, Cho, ohjoon kwon, Patrik Lambert, Mathias Buus, Amril NurmanAnnouncement
BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
Summary
Tether AI Research introduces BotPsy, a 0.8B-parameter model designed for on-device function calling and agentic workflows, built on Qwen3.5-0.8B. It achieves state-of-the-art tool-use and agentic performance for its size, outperforming every model up to 1.2B parameters on five of six benchmarks, including beating Qwen3.5-2B-Think on BFCL v4, ACEBench, τ-bench, ToolSandbox, and BrowseComp-Plus. BotPsy roughly doubles the agentic capability of its base model, with BFCL v4 rising from 23.4 to 44.0 and τ-bench from 23.2 to 51.5. It demonstrates improved safety, with unsafe responses dropping from 36% to 21% in automated red-teaming, and more efficient reasoning, using about 3× fewer reasoning tokens on GPQA Diamond than its base model. The model is released with open weights under Apache 2.0 for research and educational use in bf16, Q8, and Q4 precisions.
Context
Frontier cloud LLMs carry high operational costs, introduce vulnerabilities from third-party dependency, and offer limited customization while posing risks to privacy and data control. Many routine tasks do not require frontier-level intelligence, and small specialized agents can be equally capable on specific tasks while being significantly cheaper to run on-device.
Approach / What changed
BotPsy is a 0.8B-parameter model built on Qwen3.5-0.8B and post-trained in four sequential stages: Supervised Fine-Tuning (SFT), Multi-teacher On-Policy Distillation (MOPD), and two rounds of Reinforcement Learning (RL).
Takeaways
BotPsy outperforms every model up to 1.2B parameters on five of six tool-use and agentic benchmarks, including beating Qwen3.5-2B-Think (2.5× its size) on BFCL v4, ACEBench, τ-bench, ToolSandbox, and BrowseComp-Plus.
BotPsy reduces unsafe responses from 36% to 21% in automated red-teaming and from 46% to 31% under prompt-injection attacks compared to its base model.
BotPsy uses about 3× fewer reasoning tokens than its base model on GPQA Diamond (3,690 vs 11,563) and reduces repetitive reasoning loops from 22.9% to 1.4%.
FAQ
What types of tasks is BotPsy suitable for?
BotPsy is best suited to tasks of low to medium difficulty, such as device and system management, personal productivity, communication, information retrieval, media and lifestyle, and commerce and services. It is not intended for safety-critical or high-stakes automated decisions.
Under what license and formats is BotPsy available?
BotPsy is released with open weights under the Apache 2.0 license for research and educational use. It is available in three precisions (bf16, Q8, Q4) as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp.
Which languages does BotPsy support?
BotPsy is primarily trained and tested in English. It has been evaluated on Multi-IF covering seven additional languages (Chinese, French, Hindi, Italian, Portuguese, Russian, Spanish) and performs well in most but scores much lower in Hindi and Russian.
Source and freshness
- Source
- huggingface.co
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.