M4 Max
苹果M4系列高端芯片，内存带宽高达546GB/s，是主流DDR5系统内存带宽的4倍以上，采用LPDDR5X专用内存，支持高性能GPU渲染和神经网络推理
Core Facts
The unified memory bandwidth of the M3 Max chip is as high as 400GB/s, far exceeding the approximately 64GB/s of a traditional PCIe bus
A Mac Studio M4 Ultra equipped with 192 GB of unified memory can theoretically run complete 70B or even larger parameter models.
The Apple M2 Ultra supports up to 192 GB of unified memory, theoretically enabling the direct running of large models with tens of billions of parameters.
Timeline (last 90 days)
On M4 Max, the text route using Omni 3B + TTS 1.7B (4.61 GB) has a median full-turn latency of 4.27 seconds, with a peak process RSS of 3.86 GiB.
On M4 Max, Omni 3B has a median dialogue generation time of 171 ms and a peak process RSS of 2.57 GiB; Qwen3.5 0.8B has a median dialogue generation time of 96 ms and a peak process RSS of 1.08 GiB; Qwen3-TTS 1.7B has a synthesis time of 1.78 seconds and a peak process RSS of 3.85 GiB.
Under high power operation, the 35B MoE model reaches 78 to 107 tokens/s, while the dense 27B model only reaches 22 tokens/s.
Based on public data for high power plus MLX or Ollama, the decoding speed of different models on M4 Max is roughly a ladder as follows: Nemotron 3 Nano about 160 tokens/s, Qwen 3.5 35B at 107 to 130 tokens/s, Gemma 4 26B about 75 tokens/s, DeepSeek V4 Flash about 27 tokens/s, and dense 27B about 22 tokens/s.
In a llama.cpp environment with Qwen3 27B, the R9700 achieved 27.1 tokens/s (approximately 65% of bandwidth ceiling) and the B70 achieved 21.4 tokens/s (just over approximately 50%).
The decode write speed improvement from M4 Max to M5 Max is only 25%, constrained by memory bandwidth.
The M4 Max of the same name comes in two bandwidth variants—410 GB/s and 546 GB/s—and the M5 Max exists in 460 and 414 versions.
When using only two machines — the Mac Studio and the M4 Max — tasks exceeding one hundred thousand tokens caused crashes due to insufficient memory, which was resolved after adding the third machine.
The blogger used an M4 Max, an M5 Max, and a Mac Studio to form a three-machine cluster to run the model.
In the time-to-first-token test on a 70-billion-parameter model, the M3 Ultra took 31.6 seconds while the M5 Ultra took only 7.7 seconds, achieving approximately 4.1x, close to Apple's claimed 4.3x.
1 more timeline events
All Facts (15)
The unified memory bandwidth of the M3 Max chip is as high as 400GB/s, far exceeding the approximately 64GB/s of a traditional PCIe bus
80%VerifiedA Mac Studio M4 Ultra equipped with 192 GB of unified memory can theoretically run complete 70B or even larger parameter models.
70%VerifiedThe Apple M2 Ultra supports up to 192 GB of unified memory, theoretically enabling the direct running of large models with tens of billions of parameters.
65%VerifiedThe M4 Max already supports 128 GB of unified memory
65%UnverifiedAn M4 Max with 128GB unified memory can directly hold large model weights
80%UnverifiedIn the time-to-first-token test on a 70-billion-parameter model, the M3 Ultra took 31.6 seconds while the M5 Ultra took only 7.7 seconds, achieving approximately 4.1x, close to Apple's claimed 4.3x.
60%UnverifiedOn M4 Max, Omni 3B has a median dialogue generation time of 171 ms and a peak process RSS of 2.57 GiB; Qwen3.5 0.8B has a median dialogue generation time of 96 ms and a peak process RSS of 1.08 GiB; Qwen3-TTS 1.7B has a synthesis time of 1.78 seconds and a peak process RSS of 3.85 GiB.
50%UnverifiedOn M4 Max, the text route using Omni 3B + TTS 1.7B (4.61 GB) has a median full-turn latency of 4.27 seconds, with a peak process RSS of 3.86 GiB.
50%UnverifiedUnder high power operation, the 35B MoE model reaches 78 to 107 tokens/s, while the dense 27B model only reaches 22 tokens/s.
50%UnverifiedBased on public data for high power plus MLX or Ollama, the decoding speed of different models on M4 Max is roughly a ladder as follows: Nemotron 3 Nano about 160 tokens/s, Qwen 3.5 35B at 107 to 130 tokens/s, Gemma 4 26B about 75 tokens/s, DeepSeek V4 Flash about 27 tokens/s, and dense 27B about 22 tokens/s.
50%UnverifiedIn a llama.cpp environment with Qwen3 27B, the R9700 achieved 27.1 tokens/s (approximately 65% of bandwidth ceiling) and the B70 achieved 21.4 tokens/s (just over approximately 50%).
50%UnverifiedThe M4 Max of the same name comes in two bandwidth variants—410 GB/s and 546 GB/s—and the M5 Max exists in 460 and 414 versions.
50%UnverifiedThe decode write speed improvement from M4 Max to M5 Max is only 25%, constrained by memory bandwidth.
50%UnverifiedWhen using only two machines — the Mac Studio and the M4 Max — tasks exceeding one hundred thousand tokens caused crashes due to insufficient memory, which was resolved after adding the third machine.
50%UnverifiedThe blogger used an M4 Max, an M5 Max, and a Mac Studio to form a three-machine cluster to run the model.
50%