Tether's BotPsy-0.8B runs agent tool calls on your phone
An 800M-parameter open model built on Qwen3.5 beats every model under 1.2B parameters on six tool-use benchmarks, entirely on-device.
Tether’s AI research team released BotPsy-0.8B, an open-weight model built specifically for agentic tool calling on-device: alarms, email, calendar and travel tasks handled without sending anything to a cloud provider.
It is built on Qwen3.5-0.8B and trained in four stages: supervised fine-tuning, distillation from three larger teacher models, then two rounds of reinforcement learning. On BFCL v4 it scores 44.0 against the base model’s 23.4 and Qwen3.5-2B’s 37.7. On ToolSandbox it scores 64.6 against 30.5 and 41.2 for the same two comparisons. Tether says it leads every model of 1.2B parameters or fewer it tested, across six agent and tool-use benchmarks. Weights are on Hugging Face under Apache 2.0.
Why it matters: this is small enough to run on a phone or laptop, so the model making decisions about your inbox or calendar never leaves the device. For anyone building a local-first agent, Apache 2.0 weights at this size and this benchmark position are a real option, not just a toy.
The caveat: Tether’s own documentation says it is built for “low to medium difficulty” tasks and is not intended for safety-critical or high-stakes automated decisions.