Tether says tiny 800M-parameter BotPsy agent beats larger phone-task models

Tether has released QVAC BotPsy, an 800-million-parameter AI agent built to carry out multi-step tasks on personal devices. In company testing, the small model scored higher than competitors with up to 2.5 times as many parameters on five of six tool-use evaluations. The result matters because it points to assistants that could manage phone settings, messaging and scheduling locally, potentially without sending personal data to cloud servers.

Small model trained for tool use and recovery

BotPsy is built on Alibaba Qwen3.5-0.8B, a small language model that Tether trained further to specialize in interacting with software tools. Unlike chatbots that only answer questions, the agent is designed to identify needed tools, execute commands in order and adjust when something fails. Tether described a test case where an assistant had to send a message with Wi-Fi disabled and a full inbox: BotPsy turned on Wi-Fi, cleared storage and completed the task, while the base model repeated the failed operation.

According to Tether technical documentation cited by Decrypt, BotPsy scored 44.0 percent on BFCL v4, which measures selection and execution of software functions. The base Qwen3.5-0.8B scored 23.4 percent and the larger Qwen3.5-2B reached 37.7 percent. On ACEBench, which evaluates accuracy of function calls including tools and arguments, BotPsy scored 58.7 percent, against 30.1 percent for its base model and 55.4 percent for the 2-billion-parameter model. Across six evaluations of five models, BotPsy had the highest result on five tests, with an average of 45.0 percent compared with 38.8 percent for the larger Qwen model.

What on-device agents mean for money and privacy

For holders of USDT and users of crypto-linked apps, the practical interest is in local execution and payments. Tether first introduced QVAC in May 2025 as an alternative to AI services run by large technology firms, with longer-term plans for agents that interact with applications and potentially make payments in Bitcoin and USDT. The project later added a 41-billion-token dataset for science and engineering models in October 2025, QVAC Workbench for running models locally, a software kit for offline AI apps on mobile and desktop, and QVAC MedPsy for medical use on consumer hardware.

These figures come from Tether own testing and do not prove independent leadership or readiness as a consumer product. The company has not announced a release date for a finished assistant based on BotPsy. BotPsy is available as an open-weight model under the Apache 2.0 license, with compressed versions for local deployment, but the training code and training data were not released. The claims therefore affect developers exploring offline agents more than ordinary users at present.