Mac Studio, M5 Max, 64GB vs Mac Studio, M3 Ultra, 96GB, refurbished
Both hold Gemma-3-27B at BF16. Mac Studio, M3 Ultra, 96GB, refurbished decodes an est. 9.9 tok/s against 7.4, and costs $830 more — $336.36 per extra token per second. Worth it if you read the output as it arrives; not if you queue jobs.
- Fit
- Tight
- Fits
- Headroom
- 2GB
- 31GB
- Est. speed
- 7.4 tok/s
- 9.9 tok/s
- Cheapest today
- $3,499 at Apple
- $4,329 at Apple Certified Refurbished
- Per GB of memory
- $55
- $45
- Memory
- 64GB unified memory
- 96GB unified memory
- Usable for a model
- 59GB
- 88GB
- Memory bandwidth
- 614GB/s
- 819GB/s
- Storage
- 512GB
- 1TB
- Lowest recorded
- $3,499 in 3d
- $4,329 in 3d
Fit, headroom and speed are computed from each vendor’s published bandwidth and the model’s own size at 8K context. Prices are the cheapest listing we could read on each seller’s page. The arithmetic is here.
Where they actually differ
- Qwen3-235B-A22BWon't fitWon't fit
- DeepSeek-V3Won't fitWon't fit
- Llama-3.3-70BWon't fitWon't fit
- gpt-oss-120bWon't fitWon't fit
- Mistral-Large-2411Won't fitWon't fit
- Llama-4-Scout-17B-16EWon't fitWon't fit
- Qwen3-32BWon't fitFits
- Gemma-3-27BTightFits
- gpt-oss-20bFitsFits
- Llama-4-Maverick-17B-128EWon't fitWon't fit
- Llama-4-BehemothWon't fitWon't fit
- Qwen3-Coder-480B-A35B-InstructWon't fitWon't fit
- Kimi-K2-InstructWon't fitWon't fit
- GLM-4.6Won't fitWon't fit
- DeepSeek-R1Won't fitWon't fit
- MiniMax-M2Won't fitWon't fit
- Qwen3-30B-A3BWon't fitFits
- Qwen3-4BFitsFits
- Devstral-Small-2507FitsFits
- Phi-4FitsFits
- Gemma-3n-E4BFitsFits
- Qwen3-VL-8B-InstructFitsFits
Every model we track at BF16. The rows where the two boxes agree are dimmed, because they are not what you are choosing between.