awesome-tiny-models

A curated list of tiny open-source models that you can actually run on a laptop, a phone, or in a browser tab — text, vision, audio and video — each with its licence, size, download and checksum stated on one page.

Every entry links to its page on tinymodels.co, where each model carries a verified sha256 for its download and one measured latency figure: one fixed input per task type, one warm-up run then one timed run, on one machine, with the method stated. The full table for the latest run is published as the measured bench dataset (dated permalinks per run, raw JSON).

Method note. The measured figures are from one laptop (Apple M4 Pro, 48 GB unified memory), one run per model. They are honest order-of-magnitude numbers, not benchmark submissions: factors of two are noise, and a whole set is always re-measured together, never one model alone, so numbers stay comparable. A figure is published or absent — never estimated.

Text generation

• SmolLM2-135M-Instruct — HuggingFaceTB/SmolLM2-135M-Instruct — a 135M-parameter instruction-tuned chat model small enough to run in a browser tab.
• Qwen2.5-0.5B-Instruct — Qwen/Qwen2.5-0.5B-Instruct — half a billion parameters of general chat and instruction following, under 1 GB in bf16.
• Qwen3-0.6B — Qwen/Qwen3-0.6B — the 0.6B Qwen3 dense model: on-device chat with a switchable thinking mode.
• Qwen2.5-0.5B-Instruct-GGUF — Qwen/Qwen2.5-0.5B-Instruct-GGUF — the same 0.5B instruct model as a 491 MB Q4_K_M GGUF, ready for llama.cpp and Ollama.
• Qwen2.5-0.5B-Instruct-4bit — mlx-community/Qwen2.5-0.5B-Instruct-4bit — 4-bit MLX conversion, 278 MB, tuned for Apple silicon.

Sentence similarity

• all-MiniLM-L6-v2 — sentence-transformers/all-MiniLM-L6-v2 — the 22M-parameter sentence embedder that quietly powers a large share of local search.

Text classification

• DistilBERT SST-2 — distilbert/distilbert-base-uncased-finetuned-sst-2-english — a 67M-parameter sentiment classifier: positive or negative, distilled from BERT.

Image classification

• ResNet-50 — microsoft/resnet-50 — 25M parameters, 1000 ImageNet classes, the baseline that still refuses to die.

Object detection

• DETR ResNet-50 — facebook/detr-resnet-50 — 42M parameters of end-to-end object detection.

Vision-language (image-text-to-text)

• Florence-2-base-ft — microsoft/Florence-2-base-ft — 232M parameters covering captioning, OCR, detection and segmentation behind one prompt format.

Video

• SmolVLM2-256M-Video-Instruct — HuggingFaceTB/SmolVLM2-256M-Video-Instruct — a 256M-parameter vision-language model that watches video, not just stills.
• VideoMAE-base (Kinetics) — MCG-NJU/videomae-base-finetuned-kinetics — VideoMAE-base at 87M parameters, labelling 400 human actions from a clip.

Contributing

Entries are curated the same way the catalogue is: the model must exist upstream, state its licence and file size, and have a page on tinymodels.co with its checksum and measured figure. PRs welcome — please keep "tiny" honest (laptop-class sizes) and keep one model per entry.