Tether Open-Sources VisionPsy-Nano, a 460-Million-Parameter On-Device Vision Model
Tether Data open-sourced VisionPsy-Nano, a 460-million-parameter vision-language model that topped industry benchmarks for on-device deployment.
Overview
Tether Data's AI research initiative, QVAC, announced the open-source release of VisionPsy-Nano on 29 July 2026. The model is a compact 460-million-parameter vision-language model (VLM) built specifically for on-device and edge deployment, bringing photo and document understanding that previously needed cloud infrastructure directly onto phones. In a post on Tether's newsroom, the company said VisionPsy-Nano sets a new state-of-the-art for compact VLMs, reaching an overall normalized score of 62.3 and the number 1 spot for quality and performance among evaluated sub-0.5B on-device vision-language models.
How it stacks up
Tether put VisionPsy-Nano-460M up against three named rivals in its size class: Liquid AI's LFM2.5-VL-450M, Hugging Face's SmolVLM2-500M, and its own predecessor, nanoVLM-460M-8k. Out of a 17-benchmark suite, the new model won 16. The scoreboard: VisionPsy-Nano-460M landed an overall normalized 62.3, well clear of LFM2.5-VL-450M's 59.6, nanoVLM-460M-8k's 54.9, and SmolVLM2-500M's 52.5.
The model also leads on the categories Tether evaluated, covering document reading, scene understanding, reasoning, and instruction following. On visual perception it leads by a +4.6% relative margin over the next-best ~0.5B model, and on reasoning and knowledge it leads by +7.4%. On instruction-following tests scored 42.3 and 87.9, VisionPsy-Nano-460M beats models 1.6x to 2.3x its size, including FastVLM-0.5B (759M), Qwen3.5-0.8B (873M), and InternVL3.5-1B (1061M).
Two variants, different priorities
Tether is releasing the model in two variants. VisionPsy-Nano-460M is tuned for best-in-class, reference accuracy across standard vision-language benchmarks at the ~460M scale. VisionPsy-Nano-460M-Flash is tuned for latency, engineered for fast real-world deployment on everyday smartphones while retaining about 99% of full model quality, a 61.4 normalized score. Because it processes far fewer visual tokens, the Flash variant produces its first output tokens dramatically faster than SmolVLM2-500M and the original base model — about 19-23x quicker on the Pixel 9, Galaxy S23, and Galaxy S25 Ultra test phones, and up to about 36x quicker on an iPhone 15.
Access for developers
Both variants ship open-weight under the Apache 2.0 license, intended for researchers and educational use. Tether lists three inference paths available on day one: full-precision quality through Hugging Face Transformers, on-device mobile deployment via quantized GGUF checkpoints designed to run through llama.cpp, and high-throughput server inference powered by vLLM. All evaluation configurations and judge setups under the VLMEvalKit protocol are included so outside researchers can reproduce the benchmark results.
Paolo Ardoino, CEO of Tether, framed the release around reducing reliance on cloud data centers. He said the technology is ready to "counter the industry's obsession with data center centralization right now." He added that the results prove "local-first, highly efficient AI is a viable pathway," and that the goal is that developers building with the open weights end up "bypassing centralized gatekeepers and making powerful, private AI possible on the devices people already own."
Key facts
- VisionPsy-Nano is a 460-million-parameter vision-language model from Tether Data's QVAC research initiative, open-sourced on 29 July 2026 under an Apache 2.0 license.
- It reached an overall normalized score of 62.3, the top result among evaluated sub-0.5B on-device VLMs, beating Liquid AI's LFM2.5-VL-450M (59.6) and Hugging Face's SmolVLM2-500M (52.5) on 16 of 17 benchmarks.
- The VisionPsy-Nano-460M-Flash variant scores 61.4, about 99% of full model quality, while reaching first-token generation roughly 19-23x faster than SmolVLM2-500M on several test phones and up to about 36x faster on an iPhone 15.
- On its instruction-following tests, scored 42.3 and 87.9, VisionPsy-Nano-460M outperforms larger models including FastVLM-0.5B (759M), Qwen3.5-0.8B (873M), and InternVL3.5-1B (1061M).
- Developers can use full-precision weights through Hugging Face Transformers, run quantized GGUF builds on-device through llama.cpp, or serve the model at scale with vLLM.
Check your understanding
- What overall normalized score did VisionPsy-Nano-460M reach, and on how many of the 17 tested benchmarks did it lead?
- Which two inference paths let a developer run VisionPsy-Nano directly on a phone rather than through a server?