Tether AI Research has released VisionPsy-Nano, a family of compact vision-language models (VLMs) designed for on-device and edge deployment. The models, which include VisionPsy-Nano-460M and VisionPsy-Nano-460M-Flash, are optimized for performance and efficiency, enabling multimodal understanding without relying on cloud infrastructure. The release comes as part of an effort to bring advanced AI capabilities to mobile devices, addressing the need for privacy, latency, and offline functionality. The models are available for public use with open weights and reproducible evaluation tools.
VisionPsy-Nano-460M leads its weight class in four key areas: document understanding & OCR, visual perception, reasoning & knowledge, and instruction following & reliability. It achieves a normalized score of 62.3 across 16 of 17 benchmarks, outperforming models like LFM2.5-VL-450M, SmolVLM2-500M, and nanoVLM-460M-8k. In the Instruction Following & Reliability category, it surpasses models up to 2.3x its size, including FastVLM-0.5B, Qwen3.5-0.8B, and InternVL3.5-1B. VisionPsy-Nano-460M-Flash, optimized for latency, reaches the first token 19–23x faster than its predecessor on devices like the Pixel 9 and Galaxy S23, while retaining 99% of the full model's quality.
The models are built on nanoVLM, a compact VLM architecture that pairs a SigLIP2 vision encoder with a SmolLM2 language backbone. Tether AI designed VisionPsy-Nano to balance quality and deployability, addressing the trade-offs between accuracy and resource constraints in on-device AI. The release includes GGUF quantized checkpoints for mobile use and vLLM for high-throughput server inference.
Source: huggingface
