Small models you can actually run
Every model here is small, and every download comes with the sha256 to check it against. The files stay with their publishers: this is a catalogue, not a copy.
-
SmolVLM2-256M-Video-Instruct
HuggingFaceTB
A 256M-parameter vision-language model that watches video, not just stills.
video-text-to-textsafetensors256M1.03 GBapache-2.0 -
Qwen3-0.6B
Qwen
The 0.6B Qwen3 dense model: on-device chat with a switchable thinking mode.
text-generationsafetensors752M1.50 GBapache-2.0 -
Florence-2-base-ft
microsoft
232M parameters covering captioning, OCR, detection and segmentation behind one prompt format.
image-text-to-textsafetensors232M463 MBmit -
SmolLM2-135M-Instruct
HuggingFaceTB
A 135M-parameter instruction-tuned chat model small enough to run in a browser tab.
text-generationsafetensors135M269 MBapache-2.0 -
Qwen2.5-0.5B-Instruct-4bit
mlx-community
4-bit MLX conversion of Qwen2.5-0.5B-Instruct, 278 MB, tuned for Apple silicon.
text-generationmlx494M278 MBapache-2.0 -
Qwen2.5-0.5B-Instruct-GGUF
Qwen
The same 0.5B instruct model as a 491 MB Q4_K_M GGUF, ready for llama.cpp and Ollama.
text-generationgguf494M491 MBapache-2.0 -
videomae-base-finetuned-kinetics
MCG-NJU
VideoMAE-base at 87M parameters, labelling 400 human actions from a clip.
video-classificationsafetensors86.5M346 MBcc-by-nc-4.0 -
detr-resnet-50
facebook
DETR with a ResNet-50 backbone: 42M parameters of end-to-end object detection.
object-detectionsafetensors41.6M167 MBapache-2.0 -
Qwen2.5-0.5B-Instruct
Qwen
Half a billion parameters of general chat and instruction following, under 1 GB in bf16.
text-generationsafetensors494M988 MBapache-2.0 -
resnet-50
microsoft
ResNet-50: 25M parameters, 1000 ImageNet classes, the baseline that still refuses to die.
image-classificationsafetensors25.6M102 MBapache-2.0 -
all-MiniLM-L6-v2
sentence-transformers
The 22M-parameter sentence embedder that quietly powers a large share of local search.
sentence-similaritysafetensors22.7M90.9 MBapache-2.0 -
distilbert-base-uncased-finetuned-sst-2-english
distilbert
A 67M-parameter sentiment classifier: positive or negative, distilled from BERT.
text-classificationsafetensors67M268 MBapache-2.0
One input per task type, named on each row, measured on Apple M4 Pro, 48 GB unified memory, macOS 15.6 with one warm-up run then one timed run. How these numbers were taken.