Tether launches 800-million-parameter on-device model BotPsy, claims tool-calling performance surpasses models 2.5 times its size.

November 11: Tether announced the launch of the open-weight model QVAC BotPsy, with 800 million parameters, based on Alibaba's open-source Qwen3.5 architecture, capable of converting everyday requests into local tool calls on users' own devices. Tether stated that in its own testing, the model set a new high in tool-calling evaluations among models under 1 billion parameters and surpassed all tested models at 1.2 billion parameters and below. On four major benchmarks—BFCL v4, ACEBench, τ-bench, and ToolSandbox—scores roughly reached twice those of the base model Qwen3.5-0.8B-Think, while also exceeding Qwen3.5-2B-Think, which has 2.5 times its parameter count. On BFCL multi-turn tool calling, it scored 46.1, approximately double the 2B model's 23.8.

Tether introduced that BotPsy can reduce the sending of personal information such as emails and calendars to centralized service providers, and can also determine when no suitable tool is currently available, avoiding fabricated calls. Tether CEO Paolo Ardoino stated that users should decide for themselves what the assistant can access and which external services it can connect to. The model weights have been published on Hugging Face.