llmfit and Ollama: choose models for your Mac

This tutorial uses llmfit to choose a local LLM before downloading its weights, then checks the estimates with Ollama. The speaker tests an M4 Pro Mac with 24GB of unified memory and compares Qwen3-30B-A3B with Qwen3-32B.

The walkthrough covers model search, fit filters, recommended quantization and estimated tokens per second. It also explains the memory figures in Fit Analysis. One estimate uses 273GB/s of memory bandwidth and a 0.55 efficiency factor, with a stated error margin of about ±30%. The quality scores are llmfit's ratings, rather than independent proof of model capability.

Before testing, the speaker recalculates for Ollama's GGUF format because llmfit defaults to MLX on the demonstrated Mac setup. With a context setting of 8192, the reported memory requirements are at least 19GB for the 30B model and 21.5GB for the 32B model.

The supplied chapter metadata identifies the faster test as Qwen3-30B-A3B: the speaker reports 39.5 tokens per second against an estimate of 39.7. The 32B test reaches only 0.12 tokens per second, with 15% of processing shifted to the CPU. These are results from this machine, not general performance guarantees. The tutorial ends with macOS Homebrew installation instructions and a reminder to keep Ollama running in the background.