Local AI Guide

See the speed before choosing hardware

Prefill and token generation speed

Qwen Flash Next on Mac Studio M5 Ultra 512GB

Simulate Qwen3.8 Flash Next on an M5 Ultra 512GB. Separate model capacity from response speed and compare with an M3 Ultra configuration with the same 512GB capacity.

Benchmark-based estimate · 4K input · one user

Last updated

103.5 ~ 184.1 tok/sDecode speed(baseline: No matching baseline)
Model fully loaded (155GB / 480GB)
First token (TTFT)
0.65 to 1.2 secondsContext 4,096 input tokens included
Prefill speed
3563.3 ~ 6334.7 tok/s0.65 to 1.1 seconds (Estimate based on published benchmarks)
Total time (E2E)
1.3 to 2.3 seconds117 output tokens included
Press Play to process the input and generate the answer.
Ready
Change device or model
1200 GB/s
Active 6B / Sangju 180B
Expand settings 4-bit · 4K · Lightning MTP
QuantizationNeed 155GB
Input context4K tokens
Lightning MTP
Baseline
Sample question:
117 output tokens

Reading this result

Even with large memory, model file size is not the whole requirement. Context and acceleration buffers also matter. This screen represents one user, not throughput from multiple concurrent agents.

Required / available memory
155.0GB / 480GB
Runtime
oMLX 0.7.0 (4d4f5a28) · mlx-lm 94cdcae · mlx 0.32.2 · native kernels · Lightning MTP on

Compare this model side by side

Using the same model on an M3 Ultra 512GB helps separate capacity from generation differences. Check the hardware page separately for the 512GB configuration's availability and ordering conditions.

Include prefillCompare generation only

Check configuration

Studio M5 Ultra 512GB · Memory 512GB

Qwen3.8-Flash-Next Q4_K_M configuration. Check that the retailer lists the same memory option.

What changes with a longer document?Expand input length, waiting time, speed and memory
Detailed analysis and all 60 hardware configurations

Popular configurations

Speed simulations by device and model

View all 20

Studio M3 Ultra 512GB × Qwen3.8-Flash-Next

50 ~ 70.3 tok/s · 3.3 to 4.5 seconds

RTX PRO 6000 96GB × Qwen3.8-Flash-Next

175.2 ~ 340.0 tok/s · 1.2 to 2.2 seconds

Qwen Flash Next on two DGX Sparks: what changes?

16.7 ~ 33.8 tok/s · 1.7 to 4.3 seconds

2× RTX 5090 64GB (PCIe) × Qwen3.8-Flash-Next

127.7 ~ 263.9 tok/s · 0.46 to 0.94 seconds