BotPsy: 0.8 Billion Parameters, Local Agent with Top Scores

Tether AI Research unveils BotPsy, an agentic model with 0.8 billion parameters designed to run on-device. It claims to achieve better results than models with up to 1.2 billion parameters, providing safer responses and reduced reasoning consumption. The weights are set to be released under the Apache 2.0 license for research and education.

Open Release Announced and Low-Stakes Use Cases

Tether AI Research plans to publish the weights of BotPsy under the Apache 2.0 license for research and educational use. The team clarifies that the model can make mistakes and is not intended for high-stakes scenarios. The device is described as designed for secure workflows executed locally, with instruction tracking presented as reliable while keeping user data on the device. In terms of security, BotPsy generates responses deemed unsafe 21% of the time, compared to 36% for its base model. Regarding efficiency, it achieves its responses with about three times fewer reasoning tokens than the original model.

Benchmarks: Clear Gaps Against Models Up to 1.2 Billion

The published metrics indicate significant gains compared to the base model Qwen3.5-0.8B-Think: on BFCL v4, the score rises from 23.4 to 44.0; on ACEBench, from 30.1 to 58.7; on τ-bench, from 23.2 to 51.5. Tether AI Research asserts that BotPsy outperforms models with 1.2 billion parameters or fewer across six benchmarks of tool usage and agentic performance. The model reportedly achieves the highest score on five of the six tests, surpassing Qwen3.5-2B-Think, which is measured as 2.5 times larger. On τ²-bench in telecommunications, BotPsy reaches 82.5, ahead of this larger model. In hallucination tests on BFCL v4, it scores 86.1, described as the highest among evaluated models. The team also reports better instruction tracking than the base model.

Targeted Functions: Tools, Multi-Domain, and Local Execution

BotPsy is based on Qwen3.5-0.8B, a model with 0.8 billion parameters designed to handle function calls, agentic workflows, and intended to operate directly on-device. It is described as an assistant equipped for simple tasks, including web research, travel and lifestyle, commerce and finance, business operations and customer relationship management, device or vehicle commands, file and document organization, calendar management, as well as emails and messaging. Its dimensions make it suitable for tasks with complexity ranging from low to medium. It achieves top-tier performance in tool usage and agentic tasks for its class, surpassing models tested up to 1.2 billion parameters, according to its authors.

Method: SFT, Multi-Teacher Distillation, and Reinforcement

The training process combines four phases and three categories of signals: demonstrations, specialized teachers, and verifiable rewards. It begins with a supervised fine-tuning phase on dialogues involving tool usage. This is followed by a policy distillation phase, where the student generates its responses and receives feedback from a specialized teacher on each token. The model is then post-trained sequentially with SFT, MOPD, and two rounds of reinforcement learning. Each step starts from the best checkpoint of the previous one.

The On-Device Bet: Costs, Dependency, and Privacy

The team argues that cloud-hosted LLMs incur volatile costs and create dependency on providers who may alter or restrict access. They also highlight limitations in customization and risks surrounding privacy, data control, and transparency. Noting that many tasks do not require cutting-edge capabilities, they advocate for the use of smaller, specialized agents executed locally for routine workloads, emphasizing a rapid improvement in efficiency per parameter. In this context, BotPsy is described as designed for secure on-device workflows, with the stated goal of remaining at the forefront of this self-hosted approach.