MLX on the M5 Max 48GB

27 MLX reports on the M5 Max 48GB

Newest first. A run on several cards, with speculative decoding, with part of the model in system RAM or with batched requests says so, since each describes a different setup.

• Phi-4 · 4-bit190.8 t/s
• Phi-4 · 4-bit62.1 t/s
• Gemma 3 · 4-bit176.2 t/s
• Gemma 3 · 4-bit31.8 t/s
• Gemma 3 · 4-bit374.4 t/s
• Gemma 3 · 4-bit66.9 t/s
• Qwen3.6 · 4-bit151.1 t/s
• Qwen 3 · 4-bit186.0 t/s
• Qwen 3 · 4-bit28.4 t/s
• Qwen 3 · 4-bit147.2 t/s
• Qwen 3 · 4-bit62.6 t/s
• Qwen 3 · 4-bit366.0 t/s
• Qwen 3 · 4-bit570.6 t/s
• Qwen 2.5 Coder · 4-bit121.8 t/s
• Qwen 2.5 Coder · 4-bit28.4 t/s
• Qwen 2.5 Coder · 4-bit60.8 t/s
• Model not identified · 4-bit28.4 t/s
• Model not identified · 4-bit41.1 t/s
• Model not identified · 4-bit75.6 t/s
• Model not identified · 4-bit120.7 t/s
• Llama 3.1 · 4-bit115.4 t/s
• Llama 3.2 · 4-bit237.9 t/s
• Llama 3.2 · 4-bit527.7 t/s
• DeepSeek R1 Distill · 4-bit121.3 t/s
• Qwen 3 · 4-bit110.6 t/s
• DeepSeek R1 Distill · 4-bit28.3 t/s
• DeepSeek R1 Distill · 4-bit60.7 t/s