Mike

Great to see compact vision models getting practical. I built a privacy-first, cross-platform web UI that runs SmolVLM2-2.2B-Instruct (vision) alongside SmolLM3-3B (text). It auto-detects CUDA/MPS/CPU, pulls models on first run, and serves a clean Gradio interface.

Vision: describe images, visual Q&A, quick OCR
Text: code generation, explanation, summarization, multilingual prompts
Local only: no API keys or cloud services

I’m actively collecting feedback: ideal image sizes, better defaults for generation params, and presets that make visual tasks smoother. If you’re testing SmolVLM* locally, I’d love your notes.

Repo: https://github.com/mikecastrodemaria/SmolLM3-M2-Interface-Multimodale

Thanks for any pointers, issues, or PRs!

commented on SmolLM3: smol, multilingual, long-context reasoner about 1 year ago

Thanks for SmolLM3! I built a small, privacy-first local web UI around SmolLM3-3B (text) + SmolVLM2-2.2B (vision) with auto hardware detection (CUDA/MPS/CPU), model auto-download, and a clean Gradio interface. It’s open source and cross-platform. Would love feedback, issues, or PRs:

GitHub: https://github.com/mikecastrodemaria/SmolLM3-M2-Interface-Multimodale

Use cases: code snippets, explanations, multilingual writing, visual Q&A + light OCR.
Any ideas for presets, UX tweaks, or additional small models welcome!