Skip to content

Step 3.5 Flash on M5 Ultra 512 GB: VRAM, speed and fit

Step 3.5 Flash at 8-bit runs on the M5 Ultra 512 GB with 256K of context at about 52 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on M5 Ultra 512 GB · 8-bit (Q8_0) · Ollama / LM Studio · Chat

GPU memory
422 GB
of 512 GB
Context
256K
auto
Speed
52 tok/s
6.6 when full
First reply
1.7 s
2K prompt

Where the memory goes

Model
197 GB
Context
96.6 GB
Overhead
636 MB
OS
128 GB
Free
89.5 GB
Memory
422 GB of 512 GB
Context
256K tokens · auto
Speed
52 tok/s · 6.6 when full
First reply
1.7 s for a 2K prompt
Running cost
$5.06/mo at home · $574/mo on a cloud 8× A40

Validation: Validated

What Step 3.5 Flash on M5 Ultra 512 GB was checked against
CheckTargetResultStatus
Fits in GPU memory384 GB usable294 GB neededPass
Headroom5% free89.5 GB free (23%)Pass
Context for chat16K or more256KPass
Writing speed15 tok/s or more52 tok/sPass
First replyunder 3 s, 2K prompt1.7 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
Step 3.5 Flash
Parameters
199B, 9.4B active
Precision
8-bit (Q8_0)
Context
256K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
M5 Ultra 512 GB
GPU can use
384 GB of 512 GB

Running cost

At home
$5.06/mo
231 W at the wall
Renting instead
$574/mo
8× A40 at $4.72/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 21, 2026.

How to run it

llama.cpp
llama-server -hf ggml-org/Step-3.5-Flash-GGUF
vLLM
vllm serve stepfun-ai/Step-3.5-Flash
MLX (Mac)
mlx_lm.chat --model mlx-community/Step-3.5-Flash-4bit

Step 3.5 Flash on similar machines

All machines: Step 3.5 Flash VRAM requirements. Everything that runs here: what can the M5 Ultra 512 GB run?

Other models that run on M5 Ultra 512 GB

Questions

Can Step 3.5 Flash run on M5 Ultra 512 GB?

Step 3.5 Flash at 8-bit runs on the M5 Ultra 512 GB with 256K of context at about 52 tokens a second.

Report VR-07KA39D, calculated Nov 22, 2026 from published specs, not benchmarked.