Speed vs intelligence
BlackwellM5 Maxsame model
Intelligence: Artificial Analysis Intelligence Index v4.3, read 25 Sep 2026. The score belongs to the base model and does not measure the quantised build we ran. Qwen3.8-27B uses its default entry (26–34 with reasoning effort). Flash-Next ran as a pruned 4-bit build (REAP-288) and as a 3-bit build (512); both are plotted at the base-model score. Six benchmarked models have no index score and are left out: the Qwen3.8 4B, 9B and 35B-A3B distills, Qwen3.6-27B, Laguna XS 2.1 and DeepSeek V4-Flash.
Higher is better
BlackwellM5 Max
Throughput with many users
Total output tokens/s across all concurrent users, at the highest B both machines ran
BlackwellM5 Max
Blackwell's lead comes mostly from vLLM with NVFP4 weights and continuous batching. With the llama.cpp setting, both machines run the same engine. M5 Max MLX figures are conservative bounds.
What each user gets as load grows
Output tokens/s per user, B = 1 → 64
BlackwellM5 Max
Reading level: about 5–10 tok/s. At B=64, every model on the M5 Max falls to that level or below.
Memory needed, and the Mac that fits it
Weights (4-bit file)KV cache for one userGPU memory macOS allows by defaultBlackwell VRAM
Footprint = weights + KV cache at the selected context + about 1 GiB of runtime buffers. By default macOS lets the GPU use about ⅔ of RAM on Macs up to 36 GB and ¾ above that; the 128 GB M5 Max we tested allowed 107.5 GiB. You can raise the limit with
sudo sysctl iogpu.wired_limit_mb. KV sizes are estimates from each model family's architecture; only Llama's come from published configs. The Mac RAM verdicts barely depend on them, except for the 70B dense model at long context.Method and sources
- Blackwell: RTX PRO 6000 Max-Q (300 W cap, PCIe Gen 3), running llama.cpp Q4_K_M and vLLM NVFP4 / FP8 / bf16. M5 Max: llama.cpp Metal Q4_K_M, MLX 4-bit through LM Studio, and mlx-vlm.
- The page defaults to the fastest engine on each machine: vLLM (NVFP4 where available) or llama.cpp on Blackwell, MLX in LM Studio or llama.cpp on the Mac. The llama.cpp setting compares the same tool and the same Q4_K_M files on both machines. MoE prompt processing is at parity only in that setting, because vLLM with NVFP4 processes prompts about twice as fast as llama.cpp on Blackwell.
- LM Studio's MLX decode figures are conservative bounds. Numbers from the first, noisy pass are excluded. Full data: llm_benchmark_unified.csv, model_memory.csv, full comparison dashboard, carousel PDF.
- Intelligence scores: artificialanalysis.ai/leaderboards/models (Intelligence Index v4.3, released 7 Sep 2026).