Mistral vs Llama for Local AI on Mac

Mistral and Llama are the two open-weight model families that run well on Apple Silicon Macs through Ollama or LM Studio. For students who want offline AI (privacy, no exam-network worries), the choice is mostly about hardware fit and answer quality.

Mistral 7B (and the newer Mixtral 8x7B mixture-of-experts) are tuned for efficiency. Llama 3.1 (Meta) ranges from 8B to 70B parameters. On an M1 Max with 16 GB RAM, both 7B models run at 30+ tokens/second. The 70B Llama with 4-bit quantization fits a 32 GB Mac at maybe 5-8 tokens/second. Quality is roughly equal at 7B; Llama 70B is markedly stronger than anything 7B at any task. For exam-time use through the Fracture overlay, the Fracture app can be configured to point at a local Ollama server (http://localhost:11434) instead of the cloud.

Install Ollama (one click). Then `ollama pull mistral` or `ollama pull llama3.1`. The CLI runs the model; Fracture can point at it.

It removes the network requirement (some proctors block external API calls). The overlay still applies for invisibility to screen capture; local vs cloud LLM is orthogonal.

Mistral 7B: ~4 GB. Llama 3.1 8B: ~5 GB. Llama 3.1 70B-q4: ~40 GB.