Section 1
Speed on the M5 Max
Across all M5 Max submissions the average generation speed is about 25 tokens per second, but that average hides an enormous spread. The fastest runs come from MLX-optimized mixture-of-experts models served through vLLM: Ornith-1.0-35B reached ~85 tok/s and Qwen3.6-35B-A3B (8-bit) sustained ~69 tok/s. These models only activate a few billion parameters per token, so they feel far quicker than their 35B total size suggests.
Dense models running as GGUF on Ollama are much slower on the same hardware — gemma4:31b managed ~7 tok/s and the higher-precision qwen3.6 dense variants landed between 5 and 20 tok/s. If responsiveness matters to you, start with an MLX MoE model before reaching for a large dense one.
- Fastest: MLX MoE models such as Ornith-1.0-35B (~85 tok/s) or Qwen3.6-35B-A3B 8-bit (~69 tok/s).
- Middle ground: qwen3-coder-next runs at ~31 tok/s despite being a 79.7B model, because it is a MoE.
- Slowest: dense GGUF models on Ollama (gemma4:31b, dense qwen3.6) at roughly 5-20 tok/s.
Section 2
Output quality, coding, and tool use
For quality-sensitive work the Qwen3.6 family is the strongest on the M5 Max. qwen3.6:27b-mlx posted the highest average quality in the data (~84), and the 35B variants stay in the high-70s to low-80s. The catch is speed: the top-quality dense runs generate only 5-16 tok/s, so they suit batch or background tasks more than fast interactive chat.
For coding and tool-heavy agent work, qwen3-coder-next is the standout. As a 79.7B MoE it combines ~31 tok/s with strong coding scores (~70) and excellent role-play and planning results, and it is the model that most directly justifies buying the 64GB configuration. If you want quality and speed together without the memory cost, the MLX MoE models (Qwen3.6-35B-A3B 8-bit, ~76 quality at ~69 tok/s) are the best compromise.
- Top quality: qwen3.6:27b-mlx (~84) or qwen3.6:35b (~79), best when speed is not the priority.
- Coding and agents: qwen3-coder-next (79.7B MoE) for the strongest all-round coding and tool-use results.
- Balanced: Qwen3.6-35B-A3B 8-bit via MLX pairs ~76 quality with ~69 tok/s.
Section 3
Unified memory and model fit
Unified memory behaves differently from discrete GPU VRAM: the same 64GB pool is shared by macOS, the model weights, and the growing context cache, so treat it as a budget rather than a target. Observed footprints from the benchmarks are a good guide — a 27B model needs roughly 20-25GB, 31-35B dense models land around 26-42GB depending on quantization, and the 79.7B qwen3-coder-next MoE occupies about 50-52GB.
That headroom is exactly why the 64GB M5 Max is worth it. The largest model that ran cleanly, qwen3-coder-next, used ~51GB and left enough room for context; by contrast a dense 70B model (deepseek-r1:70b) failed to complete, effectively over budget once the OS and cache overhead are included. Aim to keep 10-14GB free for macOS and KV-cache growth rather than filling the pool completely.
- Small headroom, high quality: 27B models (~20-25GB) leave plenty of room for long context.
- The 64GB payoff: qwen3-coder-next (~51GB) runs fully on the GPU where a smaller Mac cannot.
- Avoid the ceiling: dense 70B models pushed past the usable budget and failed to complete.
Section 4
Best model by use case
If you want the shortest answer, use the balanced pick. For most 64GB M5 Max owners an MLX MoE model keeps quality high enough to be useful while staying fast enough for real interactive work, and you can step up to the large coding MoE when a task demands it.
- Balanced default: Qwen3.6-35B-A3B 8-bit on MLX — ~69 tok/s with ~76 quality.
- Fastest: Ornith-1.0-35B on MLX at ~85 tok/s when latency matters most.
- Coding and tool use: qwen3-coder-next (79.7B MoE) to spend the 64GB headroom on capability.
- Highest quality: qwen3.6:27b-mlx (~84) when you can accept 5-16 tok/s.