Back to news
launchTether AI ResearchAlibaba2026-11-11

Tether AI Research open-sources BotPsy, a 0.8B parameter tool-calling model for on-device deployment

Tether AI Research opened weights for the 0.8B parameter BotPsy model on Hugging Face on November 11, fine-tuned from Qwen3.5-0.8B and focused on function calling and agentic workflows, with GGUF quantized versions provided.

Tether AI Research open-sourced the weights of the 0.8B parameter model BotPsy on Hugging Face on November 11. The model is fine-tuned from Qwen3.5-0.8B and targets function calling and agentic workflows with on-device deployment. The model card describes BotPsy as a personal assistant that uses tools to complete various simple tasks for the user, spanning web search and research, travel and lifestyle, commerce and finance, business and customer operations, device and vehicle control, files and documents, calendar, and email and messaging.

The official note says that given the small parameter count, the model is best suited to tasks of low to medium difficulty. On benchmarks, BotPsy roughly doubles the agentic capability of its base model, with BFCL v4 rising from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. Among models at or below 1.2B parameters that the company tested, BotPsy leads on all six tool-use and agentic benchmarks, including BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval and BrowseComp-Plus, and overall beats Qwen3.5-2B-Think, a model 2.5 times its size. On knowing when not to act, it scores 86.1 on BFCL v4 hallucination tests, the highest of all models tested. Instruction following improves sharply, with IFEval rising from 59.7 to 90.8 and IFBench from 22.0 to 55.9, while general knowledge also improves, with GPQA Diamond from 33.2 to 38.1 and MMLU-Pro from 41.6 to 48.9. The model is released under Apache 2.0 with three GGUF quantized versions provided for llama.cpp and the QVAC SDK.

BotPsyTether工具调用端侧推理开源权重Qwen3.5