Smar NetworkSmar.inSmargrid.comNewsInsuranceRoomsGamesMeasurements

SMAR Daily Industry News

Daily news across AI, technology, finance, insurance, healthcare, automotive, real estate, and government

Facebook Telegram
← Back to AI News

BotPsy: On-Device Assistant Beats Larger Models

📅 2026-11-11 | 📰 Hugging Face Blog
#botpsy #on-device ai #agentic language model #tool use #model safety

Tether AI Research has released BotPsy, a 0.8 billion-parameter language model built on Qwen3.5-0.8B that is optimized for on-device personal-assistant tasks. The model is designed for function calling and agentic workflows across domains such as web search, travel, commerce, finance, business operations, device control, calendar management, and email. Because it runs locally, BotPsy aims to lower operational costs, avoid reliance on third-party cloud APIs, and keep user data private.

Benchmark testing shows BotPsy roughly doubles the agentic capability of its base model, Qwen3.5-0.8B-Think. Scores on key evaluations rise from 23.4 to 44.0 on BFCL v4, from 30.1 to 58.7 on ACEBench, and from 23.2 to 51.5 on τ-bench (average of retail, telecom, and banking subsets). Across six tool-use and agentic benchmarks—BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus—BotPsy outperforms every model tested in the ≤1.2 B-parameter class.

Even against larger peers, BotPsy leads on five of six benchmarks. It scores 44.0 versus 37.7 for Qwen3.5-2B-Think 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. The only exception is Claw-Eval, where Qwen3.5-2B-Think records 48.9 against BotPsy’s 46.5.

In multi-turn scenarios, BotPsy shows strong reliability. On τ²-bench, which requires sustained conversation, policy adherence, and action on the user’s behalf, it attains a telecom score of 82.5, surpassing Qwen3.5-2B-Think’s 69.3. Its multi-step performance leads in commerce and services (58.3 versus 47.4) and in device/system management (46.1 versus 31.9). The model also demonstrates better judgment about when to invoke tools, achieving the highest hallucination-test score of 86.1 on BFCL v4 and improving from 13.7 to 51.3 on ACEBench’s special-case tests.

Safety metrics improve alongside capability. Automated red-team testing across six harm categories shows BotPsy produces unsafe responses 21 % of the time, the lowest rate among models up to 1.2 B parameters (the base model registers 36 %). Under prompt-injection attacks, the unsafe rate drops from 46 % for the base model to 31 % for BotPsy. Reasoning efficiency also rises: the model uses about three times fewer tokens on GPQA Diamond and 2.7 times fewer on MMLU-Pro, and its loop-stuck rate falls from 22.9 % to 1.4 %.

BotPsy will be released under the Apache 2.0 license with open weights, available in bf16, Q8, and Q4 precisions as Hugging Face safetensors and GGUF files for the QVAC SDK. The release targets research and educational use, and the developers note that the model is not intended for high-stakes applications and that users should consult the Intended Use and Limitations documentation.

Why it matters

A performant, privacy-preserving on-device assistant lets enterprises cut cloud costs while keeping data under control.

FAQ

What is the parameter count of BotPsy?
BotPsy is a 0.8 billion-parameter model.

How does BotPsy’s safety compare to its base model?
In automated red-team tests BotPsy produces unsafe responses 21 % of the time versus 36 % for its base model.

Which benchmarks does BotPsy outperform larger models on?
BotPsy beats Qwen3.5-2B-Think on BFCL v4, ACEBench, τ-bench, ToolSandbox, and BrowseComp-Plus.

Is BotPsy available for commercial use?
The model is released under the Apache 2.0 license for research and educational use.

Source: Hugging Face Blog. See the original source for full details.