BotPsy: Sub-Billion Parameter Agentic Model Targets On-Device Personal Assistance

Tether AI Research has introduced BotPsy, a compact 0.8-billion-parameter language model built atop Qwen3.5-0.8B, engineered specifically for on-device agentic tasks such as web search, calendar management, and device control. By focusing on function calling and multi-turn instruction following, the model aims to deliver practical personal-assistant capabilities without relying on cloud APIs.

Benchmark Leadership in Its Weight Class

Across six standard tool-use and agentic benchmarks, BotPsy-0.8B surpasses every tested model up to 1.2B parameters, including the 2.5× larger Qwen3.5-2B-Think on five of six tests. BFCL v4 scores jump from 23.4 to 44.0, while ACEBench rises from 30.1 to 58.7, roughly doubling the base model’s agentic performance.

The model ships in safetensors and quantized GGUF formats (bf16, Q8, Q4), enabling deployment on smartphones, laptops, and embedded boards with minimal memory footprint. This on-device approach keeps user data local, eliminates API latency, and insulates applications from third-party service changes or pricing shifts.

Why It Matters for GPU / AI Infrastructure

The rise of sub-billion-parameter agents signals a growing workload segment that can run entirely on edge silicon, reducing inference demand on centralized GPU clusters. However, training and fine-tuning such specialized models still requires high-performance GPU infrastructure, and the proliferation of edge-optimized checkpoints expands the total addressable market for accelerated compute across cloud and device fleets.

BotPsy’s release underscores a broader industry trajectory: as intelligence per parameter improves, organizations can match specific workloads to right-sized models, optimizing both cost and privacy. For infrastructure providers, this means supporting a wider spectrum of model sizes and deployment targets, from massive training runs to ultra-efficient edge inference.