Tether releases BotPsy, 0.8B-parameter on-device assistant

BotPsy is Tether AI Research's answer to the cost and control problems of frontier cloud models. The Hugging Face blog post argues that frontier cloud LLMs carry high operational costs that can fluctuate unexpectedly, and that relying on third-party providers introduces vulnerabilities since access can be modified, denied, or deprecated at any time. Proprietary models also offer limited customization and pose risks to privacy and data control. The release targets that gap with a small model that is cheap to run and keeps user data on-device.

To build it, Tether started from Qwen3.5-0.8B and applied a four-stage training recipe. Supervised fine-tuning on tool-use conversations boosted tool use but pushed knowledge and instruction following below the base model. A distillation stage with three specialist Qwen3.5-4B teachers recovered those capabilities, and two reinforcement-learning rounds added further gains. The result is a model that roughly doubles its base's agentic scores and leads five of six tool-use benchmarks against models up to 1.2B parameters, including Qwen3.5-2B-Think, which is 2.5× its size.

The blog post is careful about limits. BotPsy still makes mistakes and is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health, and other sensitive topics. It is best suited to tasks of low to medium difficulty and is primarily trained and tested in English. Future development by Tether AI Research will focus on expanding on-device agentic capabilities to more hardware and local services, including decentralized peer-to-peer queries and local fine-tuning.