Local LLM model page

Moondream 2

Tiny vision model. Surprisingly good at describing images and OCR. Runs on anything.

Find the best model for my hardware Browse all 183 LLMs

Parameters

1.9B

Minimum RAM

4 GB

Model size

1.9 GB

Quantization

fp16

Can Moondream 2 run locally?

Moondream 2 is best suited for entry-level laptops and desktops. LocalClaw recommends fp16 as the default quantization, with at least 4 GB RAM.

Search term for LM Studio or compatible runtimes: moondream2

Hugging Face repository: vikhyatk/moondream2

visionlightspeed

Strengths

Tiny vision model — runs on anything
Good OCR
Great for image captioning
Apache 2.0

Limitations

Very limited context
Weak at complex reasoning
English-only

Best use cases

Image description
OCR
Visual Q&A
Document scanning
Accessibility tools

Benchmarks

Speed: 10/10

Quality: 4/10

Coding: 2/10

Reasoning: 3/10

Technical details

Developer: Vikhyat Korrapati

License: Apache 2.0

Context window: 2,048 tokens

Architecture: SigLIP vision encoder + Phi-1.5 language model

Released: 2024-07