BotPsy: 0.8B On-Device Agent Model Outperforms 2B Models

Executive Overview

Tether AI Research (Tether Data, S.A. de C.V.) has officially announced BotPsy, a 0.8 billion parameter agentic language model built on Qwen3.5-0.8B, purpose-built for function calling and agentic workflows on local devices. The model is released with open weights under the Apache 2.0 license for research and educational use, in three quantizations (bf16, Q8, Q4) in safetensors format on Hugging Face and GGUF format compatible with QVAC SDK and llama.cpp. BotPsy doubles its base model's function-calling capabilities, raising BFCL v4 from 23.4 to 44.0 and τ-bench from 23.2 to 51.5, and outperforms Qwen3.5-2B-Think on five of six benchmarks. This release signals a significant shift toward capable, private, and cost-free on-device AI agents.

Deep-Dive Features & Architecture

BotPsy aims to bridge the cost and privacy gap in frontier cloud models. The official statement notes that reliance on third-party models carries risks of access modification or revocation, plus volatile operational costs. BotPsy addresses this with a 0.8B parameter architecture that runs on-device, trained in four sequential stages: supervised fine-tuning (SFT), then multi-teacher on-policy distillation (MOPD) with three specialized Qwen3.5-4B models, then two rounds of reinforcement learning (RL) using the DAPO algorithm with rule-based rewards. Use cases include web search, travel, commerce and finance, business operations, device and vehicle control, files, calendar, and email/messaging.

On efficiency, BotPsy reaches answers with roughly 3× fewer reasoning tokens on GPQA Diamond (3,690 vs. 11,563) and 2.7× fewer on MMLU-Pro (1,912 vs. 5,105). The rate of falling into repetitive reasoning loops drops from 22.9% to 1.4%, and turns hitting the 32k token limit fall from 28.3% to 0.2%, reducing on-device response latency.

Benchmark & Competitive Performance

BotPsy outperforms all tested models ≤1.2B across six benchmarks: BFCL v4, ACEBench, τ-bench, ToolSandbox, Claw-Eval, and BrowseComp-Plus. Compared to Qwen3.5-2B-Think (2.5× larger), it leads on BFCL v4 (44.0 vs. 37.7), ACEBench (58.7 vs. 55.4), τ-bench (51.5 vs. 46.5), ToolSandbox (64.6 vs. 41.2), and BrowseComp-Plus (4.5 vs. 2.8), losing only on Claw-Eval (46.5 vs. 48.9). On τ²-bench for telecommunications it scores 82.5 vs. 69.3 for Qwen3.5-2B-Think, and on multi-turn tasks in BFCL v4 it rises from 7.5 to 46.1. It also scores 86.1 on BFCL v4 hallucination tests (highest) and 51.3 on ACEBench special cases vs. 13.7 for the base model.

Industry Impact & Enterprise Adoption

BotPsy's release represents a pivotal moment for on-device AI. By delivering competitive agentic performance in a sub-billion parameter package, it enables enterprises to deploy private, low-latency AI agents without cloud dependency or recurring token costs. The Apache 2.0 license and GGUF compatibility with llama.cpp and QVAC SDK lower barriers for integration into edge devices, mobile apps, and IoT systems. This could accelerate adoption in privacy-sensitive sectors such as healthcare, finance, and personal assistants, where data sovereignty and offline operation are critical.

Conclusion

BotPsy demonstrates that small, open, on-device models can rival larger cloud-based counterparts in function calling and agentic tasks. With its strong benchmark gains, enhanced safety profile, and efficient reasoning, it sets a new standard for accessible, private AI. As Tether AI Research continues to refine the model, BotPsy is poised to empower developers and organizations to build sophisticated local agents without compromising cost or control.

Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis