Step3 — Hardware Requirements & GPU Compatibility
VisionFunctions
Step3 is StepFun's flagship multimodal reasoning model, a mixture-of-experts vision-language system with 321 billion total parameters (316 billion in the language backbone) and about 38 billion active per token, routed across 48 experts with 3 selected plus 1 shared expert per token. It combines two co-designed techniques, Multi-Matrix Factorization Attention and Attention-FFN Disaggregation, specifically to cut decoding cost so the model stays efficient on both flagship and lower-end accelerators despite its size. Even with that efficiency focus, its total footprint still calls for a multi-GPU server rather than a single workstation card. Context length is 65,536 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in July 2025.
Step3 runs on a 256 GB+ Mac or a datacenter GPU
Best local pick → Mac Studio (M3 Ultra, 256GB) · Q4_K_M · 219.1 GB · ~16 tok/s
Macs from 192 GB (Q3_K_M · ~19 tok/s)
Apache 2.0
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Step3
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
vllm serve stepfun-ai/step3 --max-model-len 16384 --tensor-parallel-size 8vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on B300 (288 GB) (68 GB headroom) · Q4_K_M
- Generation speed
- ~113 tok/s
- generation speed
- Cost per 1M output tokens
- $17.12
- per 1M output tokens
How Much VRAM Does Step3 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 150.1 GB | 165.5 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Runs on B200 192 GB | |
| 172.1 GB | 187.6 GB | 3-bit medium — 3-bit medium quantization | Runs on B200 192 GB | |
| 211.8 GB | 227.3 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Runs on Instinct MI325X 256 GB | |
| 251.6 GB | 267 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Runs on Instinct MI350X 288 GB | |
| 291.3 GB | 306.7 GB | 6-bit, very good — 6-bit quantization, very good quality | Runs on 5× H100 SXM 80 GB | |
| 353.1 GB | 368.5 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Runs on 6× H100 SXM 80 GB | |
| 706.1 GB | 721.6 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on 6× H200 SXM 141 GB |
Selected Q4_K_M
4-bit medium quantization219.1GBat 16K context
- Weights
- 210.5 GB
- KV cache
- 8.3 GB
- Overhead
- 0.3 GB
Server-class: ~219.1 GB at 16K context · one Instinct MI325X 256 GB
Or .
Which Hardware Can Run Step3?
Q4_K_M · 211.8 GBStep3 (Q4_K_M) requires 211.8 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 276+ GB is recommended. No consumer GPU has enough memory. 2 devices with unified memory can run Step3, including Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 256GB).
192+ GB
Q4_K_M · best fit Q3_K_M (179.3 GB)
6 GPUs · 1 Mac class · 2 device families
B200~128Simulate output on B200B300~128Simulate output on B300Instinct MI350X~59Simulate output on Instinct MI350XM3 Ultra256–512 GB~19Simulate output on Mac Studio (M3 Ultra, 256GB)8x H100 SXM5~213Simulate output on NVIDIA DGX H1008x A100 SXM4~180Simulate output on NVIDIA DGX A100 640GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Consumer GPUs (up to 32 GB) and Macs up to 128 GB: too small for this model (needs ≥ 157.3 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
How good is Step3?
llmrun Score blends public benchmark boards into one number; 100 = the average open model.
Not enough data yet for an llmrun Score.
A Score needs results on enough independent boards. How the Composite Score is built
Benchmarks
This is the single board showing results for this model; badges show who ran the test.
- LMArena Text
LMArena · Oct 2026Arena
Arena1349.0
#78/188 open
Alternatives to Step3
Step 3 family · 6 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| GLM 5.3 FlashZ.ai | 124.1 | 205.2 GB | —needs > 24 GB | same size |
| IQuest Q1IQuestLab | — | 195.4 GB | —needs > 24 GB | same size |
| MiMo V2.6 Flash MOPDXiaomi | — | 186.5 GB | —needs > 24 GB | same size |
| DeepSeek V4 Flash Vision ExpDeepSeek | — | 183.3 GB | —needs > 24 GB | same size |
| Hy3Tencent | — | 182.3 GB | —needs > 24 GB | same size |
| Motif 3Motif-Technologies | — | 632.8 GB | —needs > 24 GB | same size |
Specifications
- Publisher
- StepFun
- Family
- Step 3
- Parameters
- 321.0B
- Active parameters
- 43.5B per token
- Architecture
- Step3VLForConditionalGeneration
- Context Length
- 65,536 tokens
- Vocabulary Size
- 128,815
- Release Date
- 2025-07-28
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Step3 need?
Step3 requires 211.8 GB of VRAM at Q4_K_M, or 706.1 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 321.0B × 4.8 bits ÷ 8 = 192.6 GB
KV Cache + Overhead ≈ 19.2 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run Step3?
No — Step3 requires at least 150.1 GB at Q2_K, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
What's the best quantization for Step3?
For Step3, Q4_K_M (211.8 GB) offers the best balance of quality and VRAM usage. Q5_K_M (251.6 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 150.1 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Step3 on a Mac?
Yes — Mac Pro M2 Ultra (192 GB) and 3 other Macs can run Step3, the smallest at Q3_K_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 211.8 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio (M3 Ultra, 256GB).
Can I run Step3 locally?
Yes — Step3 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Step3?
At Q4_K_M, Step3 can reach ~55 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B300 → 28.7 GB active ÷ (8000 × 0.65) = 5.52 ms, plus 61 layers × 0.055 ms = 3.35 ms, so 1000 ÷ 8.87 ms = ~113 tok/s
Estimated speed at Q4_K_M (211.8 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Step3?
At Q4_K_M, the download is about 192.58 GB. The full-precision BF16 version is 641.94 GB. The smallest option (Q2_K) is 136.41 GB.
Which GPUs can run Step3?
No single consumer GPU has enough VRAM to run Step3 at Q4_K_M (211.8 GB). Multi-GPU or professional hardware is required.
Which devices can run Step3?
2 devices with unified memory can run Step3 at Q4_K_M (211.8 GB), including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.