Tether released BotPsy, a 0.8B on-device agent that beats a 2B model

A personal assistant that runs on the phone itself has no cloud bill and no rate limit. No provider can pull access. Tether AI Research released BotPsy under Apache 2.0.

BotPsy is a 0.8 billion parameter agentic language model. Tether posted it on Hugging Face. It is built on Qwen3.5-0.8B. Eleven researchers wrote the paper, including Milan Gritta, Jihye Back and Mathias Buus from Tether Data.

What the numbers say

The point of a model this small is tool use. BotPsy is trained to pick a function, call it, read the result and decide what to do next. A phone can then drive a calendar, a browser or a car console on its own.

Tether reports BotPsy scores 44.0 on BFCL v4, a standard tool-use benchmark. Its base model, Qwen3.5-0.8B-Think, scores 23.4.

On the τ²-bench telecom subset, Tether says BotPsy scores 82.5. Qwen3.5-2B-Think scores 69.3 on the same test. That 2B model is 2.5 times the size of BotPsy. Telecom tests how a model holds a multi-turn conversation.

Tether also says the model uses about three times fewer reasoning tokens than the base model. The accuracy is the same. On the phone, those tokens are where most of the cost sits. Every figure is Tether's own, from Tether's own evaluation.

The pitch is cost and control

The authors state the case directly in their first paragraph. They say frontier cloud models carry high operational costs that can fluctuate unexpectedly. They add that a provider can modify, deny, or deprecate access at any time.

The second half of the argument is privacy. A model running on the phone sees only what the owner shows it. Nobody else can revoke the account or raise the price later.

The bill drops to electricity for the easy turns

The model comes in three formats. Hugging Face safetensors in bf16. GGUF files in bf16, Q8 and Q4 for the QVAC SDK and llama.cpp.

A team can load it on an iPhone or a laptop. Point it at a few tools. Watch it try the work the product currently sends to Claude or ChatGPT. The cost of each call drops to the electricity to run it. The hard turns, where a small model gives up, go to a frontier model as before. That is the fallback pattern the paper describes.

A buyer can test the question on their own data. Which of today's agent turns would a phone-sized model finish? Which still need the cloud model? The answer sets the real monthly bill. No benchmark substitutes for that test.

Source

Tether AI Research: BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model, by Milan Gritta, Jihye Back, JinUk Cho and co-authors.