Prefill and token generation speed

How fast will local AI run on my device?

See which local LLMs fit on your Mac or GPU and how fast they run. Compare prefill, token generation, memory, power, price, and TCO under consistent conditions.

New to local LLMs? Start with setup and PC requirements →

Device307 GB/s

Model27B

14.7 ~ 17.1 tok/sDecode speed

Model fully loaded (17.56GB / 56GB)

First token (TTFT)Context 4,096 input tokens included

8.6 to 10.0 secondsContext 4,096 input tokens included

Prefill speed8.6 to 9.9 seconds (Estimate from user-submitted records)

412 ~ 479 tok/s8.6 to 9.9 seconds (Estimate from user-submitted records)

Total time (E2E)117 output tokens included

Approximately 15.4 to 17.9 seconds117 output tokens included

Press Play to process the input and generate the answer.

Skip prefill

Ready

Expand settings 4-bit · 4K · Baseline decode

QuantizationNeed 17.56GB

Input context4K tokens

Decode optimizationBaseline

Prefill optimizationBaseline

Sample question:

117 output tokens

Check configuration

MBP M5 Pro 64GB · Memory 64GB

Qwen3.8 27B Q4_K_M configuration. Check that the retailer lists the same memory option.

Already own it? Open the setup guide

Would another device feel faster?

Compare the same model including prefill, or just token generation.

What changes with a longer document?Expand input length, waiting time, speed and memory

Detailed analysis and all 45 hardware configurations▼

Popular configurations

Speed simulations by device and model

View all 20

Qwen3.8 27B on Mac mini M4 32GB: response speed

5.1 ~ 5.7 tok/s · 16.4 ~ 35.5 s

Qwen3.8 27B: is a Mac mini M4 Pro 48GB worth it?

12.3 ~ 13.7 tok/s · 7.8 ~ 14.3 s

Qwen3.8 27B on Mac Studio M4 Max 128GB

27.9 ~ 60.4 tok/s · 5.8 to 10.5 seconds

Qwen3.8 27B on RTX 3090: do you need an upgrade?

42.3 ~ 46.9 tok/s · 4.4 to 7.5 seconds