Skip to content

Step 3.5 Flash on M5 Max 128 GB: VRAM, speed and fit

Step 3.5 Flash at 2-bit runs on the M5 Max 128 GB with 52K of context at about 72 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on M5 Max 128 GB · 2-bit (Q2_K) · Ollama / LM Studio · Chat

GPU memory
122 GB
of 128 GB
Context
52K
auto
Speed
72 tok/s
16 when full
First reply
2.8 s
2K prompt

Where the memory goes

Model
69.6 GB
Context
20.1 GB
Overhead
636 MB
OS
32.0 GB
Free
5.7 GB
Memory
122 GB of 128 GB
Context
52K tokens · auto
Speed
72 tok/s · 16 when full
First reply
2.8 s for a 2K prompt
Running cost
$2.63/mo at home · $131/mo on a cloud 4× RTX A5000

Validation: Validated

What Step 3.5 Flash on M5 Max 128 GB was checked against
CheckTargetResultStatus
Fits in GPU memory96.0 GB usable90.3 GB neededPass
Headroom5% free5.7 GB free (6%)Pass
Context for chat16K or more52KPass
Writing speed15 tok/s or more72 tok/sPass
First replyunder 3 s, 2K prompt2.8 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
2-bit (Q2_K)
Context
52K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
M5 Max 128 GB
GPU can use
96.0 GB of 128 GB

Worth trying

  • 8-bit cache: Room for 96K context

Running cost

At home
$2.63/mo
120 W at the wall
Renting instead
$131/mo
4× RTX A5000 at $1.08/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 Max 128 GB run?

Other models that run on M5 Max 128 GB

Questions

Can Step 3.5 Flash run on M5 Max 128 GB?

Step 3.5 Flash at 2-bit runs on the M5 Max 128 GB with 52K of context at about 72 tokens a second.

Report VR-1ASJICF, calculated Nov 22, 2026 from published specs, not benchmarked.