Tether AI Research has presented BotPsy, a 0.8-billion-parameter model built to call tools and handle simple tasks for the user directly on the device. The model is built on Qwen3.5-0.8B and works as a personal assistant: it can search the web, manage the calendar, read and write emails and messages, control device settings and work with files and documents, without sending personal data to an external server.
The bet runs against a common idea, that agents must rely on giant cloud models. Tether instead bets on a small model, light enough to run locally, and measures it on six benchmarks for tool use and agents. On these tests BotPsy beats every model of up to 1.2 billion parameters the company evaluated, and on tool-use tasks it doubles the performance of its base model: on BFCL v4 it rises from 23.4 to 44.0, on ACEBench from 30.1 to 58.7 and on τ-bench from 23.2 to 51.5.

Image: blog — https://huggingface.co/blog
The most telling comparison involves the larger model in the same family. Qwen3.5-2B-Think is about two and a half times bigger than BotPsy, yet on five of the six benchmarks where they were compared BotPsy scores highest, including ToolSandbox with 64.6 against 41.2. The sixth test, Claw-Eval, goes to the larger model, with 48.9 against 46.5. In multi-turn conversations, where the agent must follow a strict policy and act on the customer's behalf, BotPsy scores 82.5 on the telecom subset of τ²-bench, against 69.3 for the 2-billion model.
On safety, in automated red-teaming across six harm categories, unsafe responses fall from 36% for the base model to 21%, the lowest rate among the models of up to 1.2 billion parameters evaluated. Under prompt-injection attacks, meaning malicious instructions hidden in tool outputs, the share drops from 46% to 31%. Tether also reports more concise reasoning: on questions all the compared models answer correctly, BotPsy uses about three times fewer reasoning tokens than the base on GPQA Diamond.
Limits remain. Being small, the model makes mistakes and is not meant for high-stakes use; its performance drops as instructions pile up and in lower-resource languages. As for distribution, Tether says the weights will be released under the Apache 2.0 license for research and educational use, in three precisions (bf16, Q8 and Q4), as Hugging Face safetensors and as GGUF files for the QVAC SDK and llama.cpp. No price has been announced, since the model is open.
Looking ahead, Tether speaks of extending on-device capabilities, making private web searches more decentralized and optimizing local fine-tuning to the individual user's preferences, with the option of delegating heavy tasks to a home PC or a trusted peer-to-peer node when the phone is connected.

No comments yet. Be the first!