# Is VisionPsy-Nano-460M SOTA?

Pending benchmarks. Announced Jul 29, 2026; no independent results yet.

As of Nov 22, 2026, VisionPsy-Nano-460M is pending benchmarks on isitsota: announced Jul 29, 2026, with no independent results on the boards we track yet.

A compact vision-language model from Tether's QVAC team for on-device use: it answers questions about a single image (documents, charts, diagrams, scene text) and is post-trained from Hugging Face's nanoVLM-460M-8k, which pairs a SigLIP2 vision encoder with the SmolLM2-360M language model. It comes in two variants, the accuracy-tuned VisionPsy-Nano-460M and VisionPsy-Nano-460M-Flash, which uses far fewer image tokens for faster responses on phones. Weights are on Hugging Face under Apache 2.0, with GGUF builds for llama.cpp; no API is offered.

Canonical: https://isitsota.com/m/visionpsy-nano-460m
Last updated: 2026-11-02
Tracking since: 2026-11-02

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | QVAC (Tether) |
| Announced | Jul 29, 2026 |
| API | Not yet · None; the weights run locally |
| Open weights | Published · qvac/VisionPsy-Nano-460M and qvac/VisionPsy-Nano-460M-Flash on Hugging Face, F32 safetensors (about 2.0 GB each), loaded in Transformers with trust_remote_code; GGUF builds (FP32 to IQ3, plus a Q8 vision projector) in the -GGUFs repos for llama.cpp |
| Parameters | About 460M per QVAC (the safetensors metadata lists 507,363,296 F32 parameters for each variant) |
| Type | Vision-language model: one image plus a text prompt in, text out; English only; multi-image prompts are outside its intended use |
| Architecture | nanoVLM design: SigLIP2 base patch16-512 vision encoder, pixel-shuffle MLP connector (64 image tokens per 512x512 tile) and the SmolLM2-360M language model; images natively at 512x512, tiled up to 2048 px |
| Base model | Post-trained from lusxvr/nanoVLM-460M-8k (MIT) in five stages: generic supervised fine-tuning, capability-targeted fine-tuning, weakness-targeted fine-tuning, TIES checkpoint merging and preference alignment with replay |
| Context | 8,192 tokens |
| Variants | VisionPsy-Nano-460M (tuned for accuracy) and VisionPsy-Nano-460M-Flash (tuned for latency): same architecture and parameter count, separate checkpoints; Flash keeps each image at its native scale before tiling, so a 512x512 image takes 64 image tokens instead of 1,088 |
| License | Apache 2.0 (model cards and LICENSE files); the nanoVLM base is MIT, and the card lists fine-tuning datasets under their own terms, including CC BY-NC-SA 4.0 (TabMWP) and CC BY-NC 4.0 (PKU-SafeRLHF-V) |
| Developer | Tether AI Research (Tether Data, S.A. de C.V.), published under the QVAC organization on Hugging Face and X |
| Evaluation setup | Lab-run VLMEvalKit harness with Qwen3.6-27B as the LLM judge; the VisionPsy support and scoring fixes were open VLMEvalKit pull requests at release |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim: VisionPsy-Nano-460M compares favorably to LFM2.5-VL-450M, SmolVLM2-500M and nanoVLM-460M-8k on 16 of 17 benchmarks (MMVet the clear exception), with a normalized overall score of 62.3 against 59.6, 52.5 and 54.9.
- Lab claim: 86.5 on ScienceQA, 42.3 on MM-IFEval and 87.9 F1 on POPE, ahead of the larger FastVLM-0.5B (759M), Qwen3.5-0.8B (873M) and InternVL3.5-1B (1061M) on those three benchmarks.
- Lab claim: VisionPsy-Nano-460M-Flash reaches the first token about 19 to 23 times faster than nanoVLM-460M-8k and SmolVLM2-500M on Pixel 9, Galaxy S23 and Galaxy S25 Ultra and up to about 36 times faster on iPhone 15 (4-bit GGUF builds), and 1.3 to 2.4 times faster than LFM2.5-VL-450M.
- Lab claim: Flash keeps about 99% of the full model's quality (normalized score 61.4 vs 62.3) with about 40% lower peak memory than nanoVLM-460M and SmolVLM2-500M.

## Sources

- LAB: [QVAC announcement on X](https://x.com/qvac/status/2082439175891284476) (2026-07-29)
- LAB: [VisionPsy-Nano (QVAC article on the Hugging Face blog)](https://huggingface.co/blog/qvac/visionpsy) (2026-07-29)
- LAB: [QVAC post on X about VisionPsy in a robot demo](https://x.com/qvac/status/2117302716808802705) (2026-11-02)
- REPO: [qvac/VisionPsy-Nano-460M on Hugging Face](https://huggingface.co/qvac/VisionPsy-Nano-460M)
- REPO: [qvac/VisionPsy-Nano-460M-Flash on Hugging Face](https://huggingface.co/qvac/VisionPsy-Nano-460M-Flash)
- REPO: [qvac/VisionPsy-Nano-460M-GGUFs on Hugging Face](https://huggingface.co/qvac/VisionPsy-Nano-460M-GGUFs)
- REPO: [qvac/VisionPsy-Nano-460M-Flash-GGUFs on Hugging Face](https://huggingface.co/qvac/VisionPsy-Nano-460M-Flash-GGUFs)
- REPO: [VisionPsy collection (QVAC on Hugging Face)](https://huggingface.co/collections/qvac/visionpsy)

Tracking continues until an independent board lists it.

## Related

- Lab: [QVAC (Tether)](https://isitsota.com/labs/qvac.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/visionpsy-nano-460m
