Ollama users ask which models fit

This is less a “best model” question than a hardware-fit question. For local LLMs, the winning move is usually to start smaller, learn the tradeoffs, then scale up once you know what your box can actually sustain.

• –Ollama’s docs make the constraint clear: bigger context windows use more memory, and systems below 24 GiB VRAM default to 4k context.
• –OpenAI says gpt-oss-20b is designed to run with 16GB of memory, which puts it squarely in the “serious but still realistic” tier for a card like this.
• –Gemma 3 spans tiny to large sizes, including 4B and 12B variants, so it’s a better playground for quick experiments and teaching than jumping straight to a huge model.
• –Quantization is the main optimization lever here: lower-bit models usually buy much better fit and speed, with a manageable quality tradeoff.
• –Ollama is the right starting layer for beginners because it hides a lot of deployment friction, but the real lesson is learning how model size, quantization, and context length interact.