BotPsy brings a 0.8B on-device agent to local AI workflows
Tether AI Research has introduced BotPsy, a 0.8B-parameter language model designed for function calling and agentic workflows on devices. Built on Qwen3.5-0.8B, the model is aimed at simple to medium-difficulty tasks while keeping user data on local hardware.
- BotPsy is a 0.8B-parameter agent built on Qwen3.5-0.8B.
- It is designed for function calling and on-device workflows.
- The model outperformed tested models of up to 1.2B parameters on six agentic benchmarks.
- It will be released under the Apache 2.0 license in bf16, Q8, and Q4 precisions.
- BotPsy is intended for low to medium difficulty tasks, not high-stakes use.
A small model for everyday assistant tasks
BotPsy is designed to use tools on a user's behalf across areas including web search and research, travel and lifestyle, commerce and finance, business and customer operations, device and vehicle control, files and documents, calendars, and email and messaging.
Tether AI Research says the model is intended for secure on-device workflows, where local processing can reduce reliance on cloud providers and help preserve data control. The company says smaller, specialized agents can handle many routine workloads without the cost or dependence associated with frontier cloud models.
Because of its parameter count, BotPsy is best suited to low to medium difficulty tasks. The developers also warn that the model can make mistakes and is not intended for high-stakes use.
Benchmark gains over similarly sized models
According to Tether AI Research, BotPsy outperformed every model of up to 1.2B parameters that the company tested across six tool-use and agentic benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus.
The model roughly doubled the agentic capability of its base model on several measures. BFCL v4 increased from 23.4 to 44.0, ACEBench rose from 30.1 to 58.7, and τ-bench improved from 23.2 to 51.5.
BotPsy also exceeded Qwen3.5-2B-Think, which the source describes as 2.5× its size, on five of the six benchmarks. It scored 44.0 versus 37.7 on BFCL v4, 58.7 versus 55.4 on ACEBench, 51.5 versus 46.5 on τ-bench, 64.6 versus 41.2 on ToolSandbox, and 4.5 versus 2.8 on BrowseComp-Plus. Qwen3.5-2B-Think led on Claw-Eval, scoring 48.9 versus 46.5.
The company also reports improvements in instruction following and general knowledge. IFEval rose from 59.7 to 90.8, while IFBench increased from 22.0 to 55.9. GPQA Diamond improved from 33.2 to 38.1, and MMLU-Pro from 41.6 to 48.9.
Safety, efficiency and availability
In automated red-teaming across six harm categories, BotPsy produced unsafe responses 21% of the time, compared with 36% for its base model, according to Tether AI Research. Under prompt-injection attacks, its unsafe-response rate fell from 46% for the base model to 31%.
The model also used fewer reasoning tokens on questions that five tested models answered correctly. On GPQA Diamond, it used 3,690 tokens compared with 11,563 for its base model. On MMLU-Pro, the figures were 1,912 and 5,105. Across IFBench attempts, repetitive reasoning loops fell from 22.9% to 1.4%, while runs reaching the 32k-token limit fell from 28.3% to 0.2%.
BotPsy will be released with open weights under the Apache 2.0 license for research and educational use. It will be offered in bf16, Q8, and Q4 precisions as Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp.
A foundation for local agents
Tether AI Research says future development will expand BotPsy's ability to work with local hardware and services. Planned directions include more robust decentralized peer-to-peer queries through Tether's Hypersearch, deeper local fine-tuning through QVAC Fabric, and delegation from smartphones to home PCs or trusted secure peer-to-peer nodes.
The company also describes plans for stronger privacy protocols and secure agent-to-agent communication across trusted local networks. For users, the immediate significance is a small, locally deployable model that can call tools and manage multi-step workflows without requiring every task to be sent to a cloud service. Its limitations mean it is positioned for routine assistance rather than high-stakes decisions.
