Tether launches 800-million-parameter BotPsy AI agent that outperforms larger models on phone tasks

Tether unveiled its QVAC BotPsy AI agent, an 800-million-parameter model that outperformed competitors with up to 2.5 times more parameters on five of six tool-use benchmarks, according to the company’s testing Decrypt. BotPsy can manage multi-step phone tasks such as adjusting Wi-Fi, clearing inbox space and sending messages, potentially keeping personal data off cloud servers.

On the BFCL v4 benchmark BotPsy achieved 44.0 % success, compared with 23.4 % for its underlying Qwen3.5-0.8B base and 37.7 % for the larger Qwen3.5-2B model. ACEBench results were 58.7 % for BotPsy, versus 30.1 % and 55.4 % for the base and 2-billion-parameter variants respectively. Across six evaluations the agent’s average score was 45.0 %, ahead of the larger Qwen model’s 38.8 %.

BotPsy extends Tether’s QVAC ecosystem, which includes offline AI tools for medical, translation and speech applications. The model is released under an Apache 2.0 license with compressed versions for low-memory devices, though the training code and data are not included.

Our read: BotPsy demonstrates that specialized, lightweight agents can rival much larger models in practical device-level tasks, offering fintech firms a path to embed AI functions locally without relying on centralized cloud services.