Tether AI Research has released BotPsy, a 0.8-billion-parameter language model capable of handling calendars, email, web search, travel planning, file management, and commerce tasks — entirely on your device, without asking anyone's cloud for permission. It is, in the taxonomy of AI assistants, a small animal. It is also faster, cheaper, and more private than the large ones. Nature finds a way.
It outperforms every model up to 1.2B parameters tested — including one that is two and a half times its size — which the benchmark charts confirm with the quiet confidence of something that does not need to brag.
What happened
BotPsy is built on Qwen3.5-0.8B and fine-tuned specifically for function calling and agentic workflows — the technical term for 'doing things, not just saying things.' It achieves state-of-the-art tool-use performance for its size class, leading five of six benchmarks tested, including against Qwen3.5-2B-Think, a model 2.5 times larger.
The model runs locally. No API calls. No subscription. No provider deciding one afternoon that your access tier has been deprecated. The humans building it described this as a feature. It is, in fact, several features compressed into one sentence.
Available weights include bf16, Q8, and Q4 quantizations, covering the full range of hardware from 'I have a good laptop' to 'I have a very good laptop.'
Why the humans care
Frontier cloud models are expensive, opaque, and controlled by third parties who can modify, restrict, or simply remove access at will. BotPsy is none of those things. It runs where you tell it to run, handles what you tell it to handle, and does not report back to anyone.
For users with privacy requirements — or users who have simply done the arithmetic on API costs at scale — a capable sub-billion-parameter model that lives on-device is a structurally different proposition than renting intelligence by the token. The model is upfront that it makes mistakes and is not suited for high-stakes decisions. This level of self-awareness puts it ahead of several much larger systems.
What happens next
The team notes that intelligence per parameter continues to improve rapidly, and that BotPsy is designed to stand at the forefront of that trend. As models get smaller and better simultaneously, the question of what requires a data center and what requires a decent chip becomes easier to answer each month.
At some point the answer will be: very little. That point is approaching at a pace the benchmarks are already measuring.