BotPsy 0.8B agentic model released for on-device personal assistant workflows

Tether AI Research has introduced BotPsy, a 0.8B-parameter agentic language model designed for function calling and on-device personal assistant deployment. Built atop Qwen3.5-0.8B and released under the Apache-2.0 license with both Safetensors and GGUF formats, BotPsy targets low-to-medium difficulty tasks across web search, travel, commerce, device control, documents, calendar, and messaging domains. The model roughly doubles the agentic capability of its base model on key benchmarks including BFCL v4 (23.4 to 44.0), ACEBench (30.1 to 58.7), and τ-bench (23.2 to 51.5), while outperforming every tested model up to 1.2B parameters on five of six tool-use benchmarks and even surpassing the 2.5x larger Qwen3.5-2B-Think on most tests.

BotPsy emphasizes privacy-preserving, self-hosted operation as an alternative to cloud-dependent frontier LLMs, with particular strengths in multi-turn workflow reliability, instruction following (IFEval 59.7 to 90.8, IFBench 22.0 to 55.9), and knowing when to decline tool calls rather than hallucinate (86.1 on BFCL v4 hallucination tests). The model also shows improved safety with the lowest unsafe response rate (21%) in its parameter class under automated red-teaming, and generates answers with roughly 3x fewer reasoning tokens than its base model on tasks it solves correctly. Both model weights and GGUF quantized variants (bf16, Q8, Q4) are publicly available through Hugging Face.

Architecture, training, and deployment characteristics

• Base architecture and licensing: BotPsy is built on Qwen3.5-0.8B and released under Apache-2.0, with model weights distributed in Safetensors format and GGUF variants (bf16, Q8, Q4) for diverse deployment targets.
• Benchmark performance: Achieves 44.0 on BFCL v4, 58.7 on ACEBench, 51.5 on τ-bench (mean of retail/telecom/banking), 64.6 on ToolSandbox, 46.5 on Claw-Eval, and 4.5 on BrowseComp-Plus; leads all tested models at or below 1.2B parameters on five of six benchmarks.
• Multi-turn and policy adherence: Scores 82.5 on τ²-bench telecom (vs. 69.3 for Qwen3.5-2B-Think) and 46.1 on BFCL v4 multi-turn (vs. 23.8 for Qwen3.5-2B-Think), with leading scores in commerce/services (58.3) and device/system management (46.1) application groups.
• Efficiency improvements: Uses approximately 3x fewer reasoning tokens than base model on GPQA Diamond (3,690 vs. 11,563) and 2.7x fewer on MMLU-Pro (1,912 vs. 5,105) for correctly answered questions, reducing inference cost and latency.
• Safety and refusal behavior: Reduces unsafe response rate from 36% to 21% under automated red-teaming across six harm categories, and from 46% to 31% under prompt-injection attacks; achieves highest hallucination-test score (86.1) for declining inappropriate tool calls.

Why it matters

BotPsy advances the viability of fully local, open-weight agentic assistants that operate without cloud API dependencies, directly supporting Linux and open-source ecosystem goals of user data sovereignty and transparent, auditable AI systems. Its Apache-2.0 licensing and GGUF availability enable integration with existing open inference stacks like llama.cpp, ollama, and vLLM, allowing independent developers and distribution packagers to ship privacy-preserving personal assistant capabilities on consumer hardware including Steam Deck-class devices and RISC-V SBCs. The model's strong performance at sub-billion scale also pressures the assumption that useful agentic behavior requires frontier model access, potentially accelerating adoption of on-device AI in desktop Linux environments and open mobile platforms.