Tether releases BotPsy to run AI agents on personal devices

A compact agent for local tasks

Tether AI Research recently released QVAC BotPsy, an open-weight model designed to function as an on-device AI agent. The goal is to take simple human requests, such as rescheduling a meeting or managing files, and convert them into specific tool calls. By running on a user's own hardware, the model aims to provide a way to interact with apps and devices without relying on a constant cloud connection.

The model is relatively small, featuring 0.8 billion parameters. This size is intended to make it efficient enough to run on consumer electronics like smartphones and laptops. Tether provides several builds, including GGUF formats, to ensure compatibility with popular local execution tools like llama.cpp and various operating systems including Android and iOS.

While small, the model is specifically optimized for 'agentic' behavior. This means it is built to understand when it needs to use an external tool to complete a task rather than just generating text. Tether positions BotPsy as a way for developers to build commerce, finance, and messaging applications that can interact with a user's local environment.

How the model was trained

To achieve high performance despite its small size, Tether used a multi-stage training process. They built BotPsy on top of Alibaba's open-weight Qwen3.5-0.8B and then applied supervised fine-tuning on conversations involving tool use. They also utilized a technique called multi-teacher on-policy distillation to transfer knowledge from larger models to this smaller student model.

During the distillation process, three different 'teacher' models—each a Qwen3.5-4B model—graded the student's responses. Each teacher was specialized in a different area: one for tool use, one for general reasoning, and one for following instructions. This allowed the smaller model to learn specific skills from larger, more capable systems while keeping its own parameter count low.

Tether also employed reinforcement learning using a system called DAPO. This system provided partial credit for near-misses, such as getting a tool name right even if the arguments were slightly off. This approach helps the model learn from mistakes that a simple pass/fail system might ignore. These methods helped BotPsy improve its scores significantly on several agent benchmarks compared to its base version.

Benchmarking and performance

Tether reports that BotPsy outperformed Alibaba's Qwen3.5-2B-Think on five out of six agent benchmarks. For example, it scored 44.0 on the BFCL v4 benchmark, while the larger Qwen model scored 37.7 on Tether's specific evaluation harness. However, the results are nuanced because the two models were tested using different settings, such as different sampling temperatures, making direct comparison difficult.

The model also showed improvements in reasoning efficiency. Tether noted that BotPsy reached correct answers on GPQA Diamond tasks using 3,690 reasoning tokens, compared to 11,563 tokens for its base model. It also significantly reduced the rate of repetitive reasoning loops, which can sometimes cause smaller models to get stuck in a cycle of repeating the same thought or action.

Despite these gains, the larger Qwen model still leads in areas requiring deep knowledge. On the GPQA Diamond benchmark, the 2B model scored 59.9 compared to BotPsy's 38.1. Tether suggests that because of this gap, BotPsy should be used to call tools for facts it does not already know. The model also performs best in English, with scores dropping significantly when tested in Hindi or Russian.

Security and safety limits

Tether includes specific warnings regarding the safety of BotPsy. The company states the model is not intended for safety-critical decisions or high-stakes automated tasks. They also explicitly advise against using it for applications involving money, personal information, or health data. These limitations are partly due to the model's current performance on sensitive tasks, such as banking benchmarks where it showed no improvement over its base version.

Security testing also revealed vulnerabilities common in small-scale agents. In automated red-teaming, BotPsy produced unsafe responses 21% of the time. More concerningly, under prompt-injection attacks—where an agent reads untrusted text like an email body—the unsafe response rate was 31%. This means nearly one in three responses remained unsafe when the model was exposed to potentially malicious external input.

Because local execution removes the cloud API, it does not automatically eliminate security risks. Security experts recommend using a default-deny network egress for self-hosted agents. This prevents an agent with tool access from opening unauthorized connections to the internet. Developers are encouraged to provide the service connections and permissions themselves to maintain control over how the model interacts with the device.

What comes next for QVAC

BotPsy is part of a broader lineup of models released by Tether under the QVAC brand. This family includes other specialized models like VisionPsy for images, TranscribePsy for speech, and MedPsy for medical contexts. The naming convention for these models is inspired by Isaac Asimov's concept of psychohistory, which Tether uses to describe its model family goals.

Tether's long-term vision involves a complex system where a phone can hand off difficult tasks to a home PC or a trusted peer-to-peer network. They plan to route web searches through encrypted peer-to-peer connections to maintain privacy. While they have released the weights for BotPsy, they have not released the specific training code or the synthetic datasets used to create them.

For now, the focus remains on providing a functional, on-device tool for developers. Users can download the weights from Hugging Face to test the model against their own tool schemas. Because the results are based on Tether's own evaluation harness, independent replication by the community will be the next step in verifying the model's true capabilities compared to larger competitors.