StepFunStep 3Step3p7ForConditionalGeneration

Step 3.7 Flash — Hardware Requirements & GPU Compatibility

VisionFunctions

Step 3.7 Flash is StepFun's 201.4-billion-parameter mixture-of-experts vision-language model, built to handle images and text together in agentic workflows such as coding assistance and search. It supports tool use and selectable reasoning intensity, trading speed for deeper reasoning. At this scale, the full weight set must be held in memory, putting it in server-class territory: running it locally needs multiple high-memory GPUs, and most people will use a hosted API instead. The model supports a 262,144 token context window, enough for long documents or extended agentic sessions. It is released under the Apache 2.0 license, allowing unrestricted commercial and research use. Published in May 2026, Step 3.7 Flash is positioned by StepFun as an efficiency-focused model for coding agents and search, adding native image understanding over the earlier Step 3.5 Flash.

201.4B params · MoE11B active262K contextApache 2.0May 202620.3K downloads462 likes85.3K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Step 3.7 Flash runs on a 192 GB+ Mac or a datacenter GPU

Best local pick → Mac Pro M2 Ultra (192 GB) · Q4_K_M · 139.2 GB · ~46 tok/s

Macs from 96 GB (IQ3_XXS · ~58 tok/s)

Apache 2.0

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Step 3.7 Flash

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on B200 (192 GB) (52 GB headroom) · Q4_K_M

Generation speed: ~258 tok/s; generation speed
Cost per 1M output tokens: $6.43; per 1M output tokens

Deployon RunPodBest optionfrom $5.98/hr

Compare GPUs →

orB300 · RunPodfrom $6.94/hr

Run Step 3.7 Flash

139.2 GB VRAM at 16K context · runs on one H200 NVL 141 GB

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/unsloth/Step-3.7-Flash-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/unsloth/Step-3.7-Flash-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/unsloth/Step-3.7-Flash-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf unsloth/Step-3.7-Flash-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/unsloth/Step-3.7-Flash-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf unsloth/Step-3.7-Flash-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1. Splits the model over 8 GPUs of one server (tensor parallel).
$vllm serve stepfun-ai/Step-3.7-Flash --max-model-len 16384 --tensor-parallel-size 8
Runs the original weights, not the Q4_K_M GGUF; needs ~443 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/Step-3.7-Flash-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/Step-3.7-Flash-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$unsloth/Step-3.7-Flash-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
B200 (192 GB)CheapestCheapest52 GB free	~258 tok/s	$6.43	Deploy on RunPodfrom $5.98/hr (opens in new tab)	192 GB · 52 GB free · ~258 tok/s · $6.43 / 1M tokens
B300 (288 GB)148 GB free	~258 tok/s	$7.46	Deploy on RunPodfrom $6.94/hr (opens in new tab)	288 GB · 148 GB free · ~258 tok/s · $7.46 / 1M tokens

Compare GPUs →

How Much VRAM Does Step 3.7 Flash Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	Runs on
60.9 GB	74.5 GB	117.9 GB	61.7 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Runs on RTX PRO 5000 Blackwell 72 GB
66.5 GB	80 GB	123.4 GB	60.9 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Runs on H100 SXM 80 GB
69.2 GB	82.8 GB	126.2 GB	62.2 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Runs on H100 SXM 80 GB
74.8 GB	88.3 GB	131.8 GB	61.8 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Runs on RTX PRO 6000 Blackwell Max-Q Workstation Edition 96 GB
85.8 GB	99.4 GB	142.8 GB	73.6 GB	3-bit · i-quant — Importance-weighted 3-bit	Runs on RTX PRO 6000 Blackwell Max-Q Workstation Edition 96 GB
91.4 GB	104.9 GB	148.4 GB	91.7 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Runs on Instinct MI250X 128 GB
94.1 GB	107.7 GB	151.1 GB	80 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Runs on Instinct MI250X 128 GB
94.1 GB	107.7 GB	151.1 GB	71.2 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Runs on Instinct MI250X 128 GB
96.9 GB	110.5 GB	153.9 GB	87.6 GB	3-bit small — 3-bit small quantization	Runs on Instinct MI250X 128 GB
99.7 GB	113.3 GB	156.7 GB	96 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Runs on Instinct MI250X 128 GB
108.0 GB	121.6 GB	165 GB	89.3 GB	3-bit medium — 3-bit medium quantization	Runs on Instinct MI250X 128 GB
110.8 GB	124.3 GB	167.7 GB	113.6 GB	4-bit · legacy — 4-bit legacy quantization	Runs on Instinct MI250X 128 GB
113.5 GB	127.1 GB	170.5 GB	102.5 GB	3-bit large — 3-bit large quantization	Runs on Instinct MI250X 128 GB
119.1 GB	132.6 GB	176.1 GB	95.3 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Runs on Instinct MI250X 128 GB
124.6 GB	138.2 GB	181.6 GB	125.3 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Runs on H200 NVL 141 GB
124.6 GB	138.2 GB	181.6 GB	114.2 GB	4-bit small — 4-bit small quantization	Runs on H200 NVL 141 GB
124.6 GB	138.2 GB	181.6 GB	97.3 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Runs on H200 NVL 141 GB
132.9 GB	146.5 GB	189.9 GB	122.1 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Runs on H200 NVL 141 GB
135.7 GB	149.2 GB	192.7 GB	122 GB	4-bit large — 4-bit large quantization	Runs on GH200 Grace Hopper Superchip 144 GB
152.3 GB	165.9 GB	209.3 GB	138.1 GB	5-bit small — 5-bit small quantization	Runs on B200 192 GB
157.8 GB	171.4 GB	214.8 GB	146.5 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Runs on B200 192 GB
160.6 GB	174.2 GB	217.6 GB	142.4 GB	5-bit large — 5-bit large quantization	Runs on B200 192 GB
182.7 GB	196.3 GB	239.7 GB	163.3 GB	6-bit, very good — 6-bit quantization, very good quality	Runs on B200 192 GB
221.5 GB	235.1 GB	278.5 GB	209.4 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Runs on Instinct MI325X 256 GB
443 GB	456.6 GB	500 GB	394 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Runs on 7× H100 SXM 80 GB

Selected Q4_K_M

4-bit medium quantization

139.2GBat 16K context

Weights: 131.7 GB
KV cache: 7.2 GB
Overhead: 0.3 GB

Server-class: ~139.2 GB at 16K context · one H200 NVL 141 GB

Or rent B200 from $5.98/hr.

Which Hardware Can Run Step 3.7 Flash?

Q4_K_M · 132.9 GB

Step 3.7 Flash (Q4_K_M) requires 132.9 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 173+ GB is recommended. No consumer GPU has enough memory. 4 devices with unified memory can run Step 3.7 Flash, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (192 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Consumer GPUs (up to 32 GB) and Macs up to 64 GB: too small for this model (needs ≥ 67.2 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →

How good is Step 3.7 Flash?

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

SciCode
Epoch AI · May 2026Lab-reported
40.1%
#31/66 open
Lab-reported
ALE-Bench
Epoch AI · May 2026Lab-reported
694.1
#22/43 open
Lab-reported
CritPt
Epoch AI · May 2026Lab-reported
2.3%
#27/66 open
Lab-reported

All 3 Step 3.7 Flash benchmarks

Alternatives to Step 3.7 Flash

Step 3 family · 6 models →

Specifications

Publisher: StepFun
Family: Step 3
Parameters: 201.4B
Active parameters: 11B per token
Architecture: Step3p7ForConditionalGeneration
Context Length: 262,144 tokens
Vocabulary Size: 128,896
Release Date: 2026-05-23
License: Apache 2.0

Resources

Frequently Asked Questions

How much VRAM does Step 3.7 Flash need?

Step 3.7 Flash requires 132.9 GB of VRAM at Q4_K_M, or 443 GB at BF16.

VRAM = Weights + KV Cache + Overhead

Weights = 201.4B × 4.8 bits ÷ 8 = 120.8 GB

KV Cache + Overhead ≈ 12.1 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

Q4_K_M

132.9 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 5090 run Step 3.7 Flash?

No — Step 3.7 Flash requires at least 60.9 GB at IQ2_XXS, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.

What's the best quantization for Step 3.7 Flash?

For Step 3.7 Flash, Q4_K_M (132.9 GB) offers the best balance of quality and VRAM usage. Q4_K_L (135.7 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 60.9 GB.

VRAM requirement by quantization

IQ2_XXS

60.9 GB~53%

IQ3_S

94.1 GB~75%

Q3_K_L

113.5 GB~86%

Q4_K_M ★

132.9 GB~89%

Q4_K_L

135.7 GB~90%

BF16

443.0 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Step 3.7 Flash on a Mac?

Yes — Mac Studio (M3 Ultra, 96GB) and 20 other Macs can run Step 3.7 Flash, the smallest at IQ3_XXS. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 132.9 GB of usable unified memory (RAM minus macOS overhead), from Mac Pro M2 Ultra (192 GB).

Can I run Step 3.7 Flash locally?

Yes — Step 3.7 Flash can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Step 3.7 Flash?

At Q4_K_M, Step 3.7 Flash can reach ~95 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 B200 → 7.3 GB active ÷ (8000 × 0.65) = 1.40 ms, plus 45 layers × 0.055 ms = 2.48 ms, so 1000 ÷ 3.88 ms = ~258 tok/s

Estimated speed at Q4_K_M (132.9 GB)

NVIDIA B200

~258 tok/s

NVIDIA B300

~258 tok/s

NVIDIA GH200 Grace Hopper Superchip

~210 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Step 3.7 Flash?

At Q4_K_M, the download is about 120.82 GB. The full-precision BF16 version is 402.73 GB. The smallest option (IQ2_XXS) is 55.38 GB.

Which GPUs can run Step 3.7 Flash?

No single consumer GPU has enough VRAM to run Step 3.7 Flash at Q4_K_M (132.9 GB). Multi-GPU or professional hardware is required.

Which devices can run Step 3.7 Flash?

4 devices with unified memory can run Step 3.7 Flash at Q4_K_M (132.9 GB), including Mac Pro M2 Ultra (192 GB), Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M2 Ultra (192 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.