Particle.news

Tether Releases 0.8B BotPsy Agent That Beats Larger Models on Tool-Use Tests

The open-weight model is aimed at on-device assistants and research, but Tether did not publish training code or a consumer assistant timeline.

Overview

  • Tether AI Research published QVAC BotPsy on Wednesday, November 11, 2026 as an open-weight 0.8‑billion-parameter model for developers.
  • In company-run tests across six tool-use evaluations BotPsy posted a mean score of 45.0 and beat a 2‑billion-parameter Qwen3.5-2B model on five of the six tasks.
  • BotPsy is built by fine-tuning Alibaba’s Qwen3.5-0.8B and is distributed under the Apache 2.0 license in GGUF formats to ease local deployment.
  • Tether positions the model for low-to-medium difficulty, multi-step agent tasks—where an agent must call functions and use apps to complete actions—and says it is suited to research and education.
  • The release advances Tether’s push for local, privacy-preserving assistants but lacks published training code and data, so independent verification and a consumer-ready product remain pending.