BotPsy: A 0.8B On-Device Agent Model for Low-Cost Automation

Tether AI Research has released BotPsy, a 0.8B-parameter language model designed specifically for on-device function calling and agentic workflows. Built on the Qwen3.5-0.8B base, BotPsy achieves state-of-the-art tool-use performance for its size, outperforming all models up to 1.2B parameters in their testing. This release represents a significant step toward running capable AI agents locally, reducing reliance on expensive and potentially unstable cloud LLM APIs.

What happened

According to the Hugging Face blog post, BotPsy is a specialized model intended for low-to-medium difficulty tasks such as web search, travel planning, commerce, and device control. The model was post-trained through four sequential stages: Supervised Fine-Tuning (SFT), Multi-teacher On-Policy Distillation (MOPD), and two rounds of Reinforcement Learning (RL).

The training process leverages three types of signals: imitation of reference trajectories, token distributions from larger teacher models, and rule-based tool-calling rewards. In the MOPD stage, the model was trained against three specialized Qwen3.5-4B teacher models to recover general instruction-following capabilities that often degrade during SFT. The RL stages used DAPO, with the first round focusing on fine-grained tool-calling rewards and the second round broadening to general-domain verifiable tasks like math and code.

Benchmark results show BotPsy roughly doubles the tool-use performance of its base model. It leads on five of six agentic benchmarks, including BFCL v4 (44.0 vs 23.4 for base), ACEBench (58.7 vs 30.1), and τ-bench (51.5 vs 23.2). Notably, it also outperforms Qwen3.5-2B-Think, a model 2.5 times its size, on these five metrics. However, Qwen3.5-2B-Think still leads on the Claw-Eval benchmark (48.9 vs 46.5). In terms of general knowledge, BotPsy improves upon its base but remains below larger models, reinforcing its design as a tool-caller rather than a knowledge store.

Why it matters for small company builders

For small teams operating on tight margins, the operational cost and security risks of frontier cloud LLMs are real constraints. Providers can change pricing, deprecate endpoints, or introduce latency spikes that break production agents. BotPsy offers a viable alternative for routine, low-stakes automation where privacy and control are paramount. Because it runs on-device, it eliminates per-token API fees for high-volume, simple tasks like parsing emails, updating calendars, or triggering basic API calls. This allows you to build autonomous workflows that are immune to external service outages. However, you must be clear about its limitations: it is not a general-purpose intelligence. It is a specialized agent that relies on tools to fetch facts it does not know. If your workflow requires complex reasoning or deep domain knowledge without external tool access, this model will fail. It is best deployed as a local orchestrator that handles the "glue" logic, calling out to larger models or specialized APIs only when necessary.

What to do this week

1. Download the BotPsy model weights from Hugging Face and deploy it on a local machine or edge device with sufficient RAM to test its inference speed and memory footprint against your current hardware.
2. Map your existing low-stakes automation workflows, such as email triage or simple data extraction, and refactor them to use BotPsy’s function-calling interface instead of a cloud LLM to measure cost and latency savings.
3. Implement a hybrid routing layer in your agent framework that defaults to BotPsy for routine tasks but automatically escalates to a larger cloud model when the task complexity score exceeds a defined threshold.

Sources

• BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model — Hugging Face Blog