Tether AI Research Unveils BotPsy, a Sub-Billion-Parameter On-Device Agent
Tether AI Research has introduced BotPsy, a 0.8B-parameter agentic language model built on Qwen3.5-0.8B designed for secure on-device deployment, tool use, and multi-turn workflows.
Frontier cloud large language models often involve unpredictable operational expenses and third-party dependencies that can pose privacy and data control challenges. Furthermore, smaller tasks do not always require massive frontier models, opening the door for efficient, specialized, self-hosted architectures. To address this, Tether AI Research has introduced BotPsy, a 0.8B-parameter agentic language model built on top of Qwen3.5-0.8B and tailored for secure on-device personal assistant workflows.
Designed primarily for low-to-medium difficulty tasks, BotPsy handles tool-use functions across domains such as web search, research, commerce, finance, business operations, device control, files, calendars, and messaging. Tether AI Research released the open weights under the Apache 2.0 license, providing safetensors and GGUF files in bf16, Q8, and Q4 precisions.
Training Methodology and Performance
BotPsy undergoes a sequential four-stage training procedure. It starts with supervised fine-tuning (SFT) on tool-use conversations, followed by multi-teacher on-policy distillation (MOPD) utilizing three specialist Qwen3.5-4B models to recover general knowledge and instruction-following capabilities. Finally, the model goes through two rounds of reinforcement learning (RL) using rule-based and verifier-computed rewards.
On benchmarks, BotPsy roughly doubles the capability of its base model, scoring 44.0 on BFCL v4, 58.7 on ACEBench, and 51.5 on τ-bench. It outperforms every tested model in the 1.2B-parameter weight class and frequently exceeds the performance of larger models like the 2.5× larger Qwen3.5-2B-Think. Additionally, BotPsy demonstrates improved reasoning efficiency, utilizing significantly fewer reasoning tokens per correct answer on datasets like GPQA Diamond and MMLU-Pro compared to its base model.
What it means for developers
For developers building local applications, BotPsy offers an efficient, privacy-focused alternative for handling multi-turn instructions, function calling, and device management without relying entirely on costly cloud infrastructure. While exploring specialized local models, developers can also try top AI models cheaply through one API at https://apixoai.online. However, because BotPsy is a small model, it is not intended for high-stakes or safety-critical automated tasks involving sensitive financial, medical, or personal information.
Limitations and Safety
Although BotPsy achieves strong performance in instruction-following—scoring 90.8 on IFEval and 55.9 on IFBench—it exhibits certain weaknesses. It struggles with open-ended search tasks, scores lower on lower-resource languages like Hindi and Russian, and remains unsuited for high-stakes automated decisions. Nevertheless, automated red-teaming shows that its unsafe response rate drops to 21%, down from 36% for its base model, while its vulnerability to prompt-injection attacks decreases to 31%.
Source: BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model — Hugging Face blog. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
