Tether AI’s BotPsy offers sub-billion-parameter agentic control for on-device workflows

Tether AI Research has introduced BotPsy, a 0.8B-parameter language model optimized for on-device agentic workflows and function calling, according to a blog post published by Hugging Face.reported The model is built upon the Qwen3.5-0.8B architecture and targets environments where cloud reliance introduces cost volatility, privacy risks, or access instability.reportedThis small AI model runs locally on devices, helping users with simple tasks like email and calendar management without sending data to the cloud.BotPsy leverages Qwen3.5-0.8B to achieve SOTA tool-use performance in the ≤1.2B class, optimizing for local inference and multi-turn consistency.

The evidence

The post states that BotPsy roughly doubles the agentic capability of its base model, Qwen3.5-0.8B-Think, across key benchmarks.reported Scores on BFCL v4 rose from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. These figures represent the mean of retail, telecom, and banking subsets for τ-bench.

BotPsy outperforms leading models in the ≤1.2B parameter class, including LFM2.5-1.2B-Think and MiniCPM5-1B-Think, on all six tested benchmarks.reported The benchmarks include BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus. This performance is attributed to enhanced instruction-following reliability and efficiency for secure on-device workflows.

In direct comparison with Qwen3.5-2B-Think, a model 2.5 times larger, BotPsy achieves higher scores on five of the six benchmarks.reported It leads on BFCL v4, ACEBench, τ-bench, ToolSandbox, and BrowseComp-Plus. The larger model only retains a lead on Claw-Eval, scoring 48.9 against BotPsy’s 46.5.reported

The decision

BotPsy demonstrates superior reliability in multi-turn workflows requiring strict policy adherence and long customer conversations. On τ²-bench, it scores 82.5 in telecom, significantly ahead of Qwen3.5-2B-Think’s 69.3. In BFCL v4’s multi-turn tasks, BotPsy’s score rises from 7.5 to 46.1, nearly double the larger model’s 23.8.reported

Grouped by application, BotPsy leads all tested models in commerce and services, scoring 58.3 versus 47.4 for the next best model.reported It also leads in device and system management, with a score of 46.1 compared to 31.9. These results suggest the model is particularly effective for structured, multi-step operational tasks.

What the evidence means

The model excels at knowing when not to act, a critical safety feature for autonomous agents. On BFCL v4’s hallucination tests, which verify that a model declines to call tools when none are appropriate, BotPsy scores 86.1.reported This is the highest score among all models tested, indicating robust restraint in tool-use scenarios.

Despite these strengths, the source material explicitly states that BotPsy is best suited for low to medium difficulty tasks. As a small model, it still makes mistakes and is not intended for high-stakes use. Developers are directed to read the Intended Use and Limitations sections for further guidance on safe deployment.

What builders can do with it

Developers can deploy BotPsy for on-device personal assistance across several specific domains. The model supports tool use for web search, travel, commerce, finance, business operations, device control, file management, calendar scheduling, and email. Its small parameter count allows for efficient processing on local hardware, reducing dependency on frontier cloud LLMs.

Builders can integrate the model using provided safetensors and GGUF formats, including bf16, Q8, and Q4 quantizations.reported This facilitates deployment on resource-constrained devices while maintaining privacy and data control. The model’s high performance in multi-turn workflows makes it suitable for customer service bots and internal business assistants that require strict policy adherence.