Tether’s 0.8B BotPsy Beats Larger Model in Five of Six AI Tests

Tether AI Research released QVAC BotPsy on Nov. 11, an open-weight model with 0.8 billion parameters that averaged 45.0 across six evaluations of AI agent tool use. It beat the two-billion-parameter Qwen3.5-2B-Think in five of those tests, a result focused on tool-use tasks rather than a broad comparison of AI capabilities.

An AI agent carries out tasks by selecting and using tools, such as functions that handle email or documents. BotPsy was built on Alibaba’s Qwen3.5-0.8B model and fine-tuned to improve its performance on agent tasks. The six evaluations included tests of whether a model can choose and use tools for assigned work.

On BFCL v4, a function-calling benchmark, BotPsy scored 44.0, compared with 23.4 for its base model. It scored 51.5 on τ-bench, while the base model scored 23.2. BotPsy also recorded a hallucination-resistance score of 86.1 on BFCL v4, the highest among competitors with fewer than 1.2 billion parameters.

The comparison with Qwen3.5-2B-Think was not a sweep: BotPsy lost one of the six head-to-head evaluations. The scores describe performance on a limited group of agent tests and do not establish how the model performs across other AI tasks.

Tether describes BotPsy as suited to low- to medium-difficulty work, including email function calls and document handling, and positions it primarily for research and education. These are separate points: email and document tasks are examples of work it may handle, while research and education describe its intended use.

The model is part of Tether’s QVAC initiative, which aims to run AI on local devices without cloud services. BotPsy was released under the Apache 2.0 license and distributed in GGUF formats designed for local hardware.