M1 Max LLM Benchmark: Larger MoE Models Prove Faster Locally

A local LLM benchmark on an Apple M1 Max with 64GB of RAM revealed that larger models are not always slower. The test, using Ollama, found that a 23.9GB Qwen3.6 MoE model achieved 60.4 tokens/sec, outperforming a smaller 6.6GB Qwen 3.5 dense model which scored 40.5 tokens/sec. This is because MoE models use fewer active parameters per token, leading to faster decoding speeds despite their larger overall size. The benchmark also highlighted the importance of avoiding cache effects during testing, especially for prefill speeds, and recommended using the median of multiple runs for reliable results. AI

IMPACT Local LLM performance insights for older hardware, particularly MoE vs. dense model behavior, can inform user choices and hardware utilization.

RANK_REASON The item details a specific benchmark and analysis of local LLM performance on older hardware, including methodology and findings. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Apple M1 Max
• Gemma4
• Ollama
• Qwen 3.5
• Qwen3.6
• WCW Uncensored

• infra
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →