VisionPsy-Nano-460M-Flash-MLX

MLX port of qvac/VisionPsy-Nano-460M-Flash — the latency-optimized sibling of VisionPsy-Nano-460M — converted to run natively on Apple Silicon.

• Architecture: SigLIP2-base-patch16-512 vision encoder + pixel-shuffle modality projector + SmolLM2-360M-Instruct decoder (same weights as Standard except the projector uses fewer visual tokens per image, ~7 tiles vs ~13 tiles at 512x512)
• Parameters: ~460M
• Precision: bfloat16 (~1.0 GB on disk, down from 2.0 GB fp32)
• Runtime: MLX on Apple Silicon (M-series)
• License: Apache-2.0

Benchmarks (MLX bf16, M-series)

Measured across 7 images x 5 prompts, 64 max new tokens, greedy decode:

Metric	Standard	Flash
Avg decode tok/s	99	152
Median decode tok/s	90	157
Avg peak GPU memory	2.64 GB	2.64 GB
Load time	~0.4 s	~0.7 s

Per-prompt-type medians (Flash):

Prompt type	Median tok/s	Example
Describe (EN, 1 sentence)	198.5	"A man in a white lab coat with the name 'De WARON' on it stands behind a desk..."
What text appears?	96.9	"De Wooning"
Count objects/people	61.4	"1"
Main subject	170.1	"The main subject is a man wearing a white lab coat with the name 'De Woon' and a logo on it."
Describe (ZH)	221.2	"这是一...話是柯尼特, 黃色胸部..."

Flash gets the biggest speedups on text-heavy images (receipts: ~295 tok/s, neon text: ~244 tok/s) where fewer visual tokens still capture the content.

Usage

Layout matches mlx-vlm conventions. Until mlx-vlm adds a visionpsy_nano handler, load via the MLX port at KaedeTai/mlx-video @ visionpsy-mlx-port:

from huggingface_hub import snapshot_download
from mlx_video.models.visionpsy_nano import load_visionpsy_nano
from mlx_video.models.visionpsy_nano.processor import load_processor
from PIL import Image

path = snapshot_download("KaedeTai/VisionPsy-Nano-460M-Flash-MLX")

model, cfg = load_visionpsy_nano(path)
proc = load_processor(path, cfg=cfg)

img = Image.open("photo.jpg").convert("RGB")
batch = proc("Describe this image in one sentence.", image=img)

tokens = list(model.generate(
    batch["input_ids"],
    pixel_values=batch["pixel_values"],
    image_token_id=batch["image_token_id"],
    max_new_tokens=64,
    eos_token_id=proc.tokenizer.eos_token_id,
))
print(proc.decode(tokens, skip_special_tokens=True))

Attribution

• Original model: Tether AI Research / QVAC — qvac/VisionPsy-Nano-460M-Flash (Apache-2.0)
• MLX port + weight repack: KaedeTai
• Base components: SigLIP2 (Google), SmolLM2 (Hugging Face)

See also

• Standard variant: KaedeTai/VisionPsy-Nano-460M-MLX
• Original release blog + benchmarks: qvac/VisionPsy-Nano-460M
• MLX port source: github.com/KaedeTai/mlx-video @ visionpsy-mlx-port