Tether AI ships BotPsy, a 0.8B on-device agent that out-tools models twice its size

Tether AI Research published BotPsy on November 11, an 0.8-billion-parameter agentic language model built on Qwen3.5-0.8B and designed for function calling and tool-use workflows on device. The team positions it as a personal assistant that calls tools to handle web search and research, travel and lifestyle, commerce and finance, device and vehicle control, files and documents, calendars, and email and messaging.

The stated motivation is cost and control. Frontier cloud models carry high and fluctuating operating costs, and relying on third-party providers means access can be changed, denied, or deprecated at any time. Proprietary models also offer limited customization for specific workflows while raising privacy, data-control, and transparency concerns.

The benchmark results are the centerpiece. Against its base model Qwen3.5-0.8B-Think, BotPsy lifts BFCL v4 from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5 across the retail, telecom, and banking subsets. In the sub-1.2B weight class it leads all six tool-use benchmarks the team tested: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus.

It also beats a model 2.5 times its size on five of those six benchmarks, including BFCL v4 at 44.0 versus 37.7 and ToolSandbox at 64.6 versus 41.2; only Claw-Eval goes the other way, 46.5 versus 48.9. On the τ²-bench telecom subset, where the agent must hold a long customer conversation and follow policy, BotPsy scores 82.5 against 69.3 for Qwen3.5-2B-Think, and its BFCL v4 multi-turn score rises from 7.5 to 46.1.

Abstention gets unusual attention. BotPsy scores 86.1 on BFCL v4 hallucination tests, the highest of the models tested, and 51.3 on ACEBench special cases involving incomplete requests, wrong parameters, or tools that cannot serve the request, up from 13.7. An agent that reads untrusted tool output needs to know when not to act, and the team treats that as a first-class capability.

Training runs in four stages: full-parameter supervised fine-tuning on tool-use conversations, then multi-teacher on-policy distillation from three RL-trained Qwen3.5-4B specialists in tool use, reasoning, and instruction following, using per-token reverse KL as the signal, and finally two rounds of reinforcement learning with rule-based tool-calling rewards.

Efficiency numbers point at on-device latency. On questions all five models answer correctly, BotPsy reaches answers with roughly 3,690 reasoning tokens on GPQA Diamond versus 11,563 for its base, and 1,912 versus 5,105 on MMLU-Pro. Repetitive-loop rates fall from 22.9% to 1.4% and runs hitting the 32k-token cap drop from 28.3% to 0.2%, while automated red-teaming across six harm categories puts unsafe responses at 21% against 36% for the base model.

BotPsy ships with open weights under Apache 2.0 for research and educational use, in bf16, Q8, and Q4, as safetensors and as GGUF files for the QVAC SDK and llama.cpp. The team is explicit about limits: it still makes mistakes, is not intended for high-stakes use, is weak on multi-step information retrieval, and lags in lower-resource languages. The roadmap points to decentralized peer-to-peer search through Hypersearch, local fine-tuning with QVAC Fabric, and phones handing heavy tasks to home PCs or trusted peers.