Benchmarks Overview¶

Quick Picks (MacBook Pro M1 Max 64 GB): See the homepage TL;DR for the full quick-picks table. In summary: Best all-around → gemma4:e4b-nvfp4 (60.4 tok/s, 8/10 quality). Best coding → qwen3-coder:30b (55.7 tok/s, 10/10 quality). Best MoE → qwen3.6:35b-a3b-nvfp4 (54.3 tok/s, 9/10 quality). Best small → gemma4:e4b-nvfp4 (60.4 tok/s, 8/10 quality). Best reasoning → deepseek-r1:14b (10/10 quality, but memory-hungry).

This section provides reproducible benchmarks for running large language models on Apple silicon. All measurements were taken on a MacBook Pro M1 Max (64 GB unified memory) running macOS 26.6 — see the Methodology page for the complete hardware configuration and measurement approach. Results scale to other M-series chips by memory bandwidth and GPU core count.

Key findings¶

• Gemma 4 delivers the best throughput-to-quality ratio on Apple silicon. The gemma4:e4b-nvfp4 variant hits 60.4 tok/s with 8/10 quality, while gemma4:26b-nvfp4 reaches 52.8 tok/s at 9/10 quality — both with excellent cold-start times under 3.5 seconds.
• Qwen 3.6 35B-A3B is the MoE sweet spot. The 35B-parameter model with 3B active parameters achieves 54.3 tok/s at 9/10 quality, rivaling dense 30B models at half the memory footprint.
• Qwen 3 Coder 30B is the top coding model at 55.7 tok/s with 10/10 quality, but requires ~15 GB KV cache at 128K context — plan memory accordingly.
• DeepSeek R1 14B is the thinking/reasoning champion at 10/10 quality, but its 40 GB KV cache at 128K context makes it a memory hog — it can only co-reside with tiny models on 64 GB systems.
• NVFP4 quantization is the default recommendation — it offers the best perplexity-to-speed tradeoff (17.95 vs BF16 baseline 17.54) and is MLX-native.
• MLX engines vary wildly in prefill performance. oMLX leads effective throughput (170.7 eff tok/s) while Rapid-MLX leads generation speed (51.4 gen tok/s) but has high prefill overhead.
• API overhead is real. /v1/chat/completions is 33–37% slower than /api/generate on the same engine.

Summary performance table¶

Architecture overview¶

The benchmark testbed runs on a MacBook Pro M1 Max (64 GB unified memory) with Ollama serving models via /api/generate. All measurements use consistent methodology — see the Methodology page for details.

Benchmark categories¶

Hardware context¶

All benchmarks were run on the test machine described in the Methodology page. See that page for the complete hardware configuration, software versions, and measurement approach.