A practical question about model choice#
The talk starts with a question: how much model do you need for the task in front of you? Its focus is small language models and the surrounding application, from local inference to multi-agent pipelines.
The original session includes demos on Apple Silicon. The resources cover local inference with MLX and Ollama, structured extraction, vision OCR, code generation, reasoning modes, browser automation and an A2A multi-agent pipeline.
What the session covers#
| Thread | What to explore |
|---|---|
| Local inference | MLX and Ollama runtimes |
| Structured extraction | NuExtract examples |
| Document understanding | PaddleOCR-VL |
| Agent workflows | A2A and browser automation |
| Model selection | Capability and deployment trade-offs |
Follow the examples#
The talk page links to the slides and the source repository. Start there for the actual examples and their setup instructions. Repository requirements and model availability can change, so use the current project documentation when running a demo.
Continue with the model directory#
slm.expert is a related project: a directory of small open-weight models organized around use cases, with published benchmarks and license information. It can help narrow a shortlist before you test models in your own application.
A few useful questions
Where are the slides and demo code?
The original talk page links to the slide deck and the gok03/slm-talk repository on GitHub.
Does an open-weight model automatically have an unrestricted license?
No. Check the license of each model and any relevant components before deciding how to use it.