ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
StepFun 阶跃星辰
Open-Weight StepFun 阶跃星辰 · MoE 604B / A27B · Text+Image+Video · BF16 / FP8

Step 5 Preview (600B MoE) Architecture Parameters, Benchmarks & VRAM

StepFun 5th-gen flagship multimodal MoE (600B total / 27B active, 92 layers) with 1M context.

🧬 Step 5 Preview (600B MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (stepfun-ai/Step-5-Preview) config.json & Model Card

HF Class: Step5ForConditionalGeneration

1. Backbone & Dimensions

step5_vl
Total / Active Params
604B / 27B
Hidden Size (d_model)
4,096
FFN Intermediate Size
13,824
MoE Expert Dim
1,536
MoE Experts
352 Routed · Top-8
Activation / Norm
SiLU (SwiGLU) · ε=0.00001
Native Tensor Dtype
BF16 / FP8

2. Attention & Layer Layout

HYBRID
Num Hidden Layers
92 Layers (23 Global KV)
Hidden Layer Layout
92-Layer Sparse MoE (352 Experts, Top-8 Active)
Attention Heads (Q / KV)
64 Q / 4 KV (16:1 GQA)
Head Dimension
192

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
1,048,576 tokens
RoPE Configuration
RoPE (原生 1,048,576 Context)
Vocab Size
128,896 (Untied)

📋 Step 5 Preview (600B MoE) Official API & System Card Specifications

Verified against official StepFun 阶跃星辰 developer documentation & pricing

API Model ID: step-5-preview

1. Identity & Capabilities

API Spec
API Model ID
step-5-preview
Release Date
2026-09-16
Weight License
Open-Weight (API Spec)
Native Reasoning
✅ Supported
Function Calling & JSON Schema
✅ 原生支持
Batch API 异步批处理
✅ Supported (50% off)

2. Context, Output & Modalities

Limits
Max Context Window
1,000,000 tokens (1000K)
Max Output Tokens
66,000 tokens
Input Modalities
Text + Image + Video
Output Modalities
Text
Direct China Access
✅ Direct CN + RMB
Deployment Mode
Open-Weight + Cloud API

3. Official Pricing & Cache

CNY / 1M
Input / 1M Tokens
¥ 7
Cached Input / 1M
¥ 0.7 -90%
Output / 1M Tokens
¥ 20
Tiered Pricing Notes
StepFun: Official StepFun CN platform

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Official Default Sampling
temperature=0.7 top_p=0.95 max_tokens=32,768

⚡ Step 5 Preview (600B MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 604B / A27B·92L (d=4096)·64Q/4KV·Ctx 1000K·Text+Image+Video
Runnable Models →
512 GB Unified·1200 GB/s·Safe Usable 460.8 GB
Smooth Fit376.18 / 460.8 GB~46.4 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
376.18/ 460.8 GB usable

Physical 512 GB · Load 81.6%

Theoretical Decode Speed
~46.4tokens/s

MoE: reads only 16.37 GB active weights/step

Max Context at Batch=1
1000K/ 1000K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 366.18 GBKV Cache (1×16K): 1.38 GBOverhead: 8.62 GB
Reserves ~10% unified memory for OS
🏃 Smooth & Snappy~10.8s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run step-5-preview-600b

Mac Studio M5 Ultra 512G Fits Step 5 Preview (600B MoE) · Cost & TCO

Approx. $9,499 (~$19/GB VRAM) · 190W TDP

📊 Step 5 Preview (600B MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 1208 GB 1.38 GB 1234.84 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 641.75 GB 1.38 GB 657.27 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 495.28 GB 1.38 GB 507.87 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 429.6 GB 1.38 GB 440.87 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 366.18 GB 1.38 GB 376.18 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 320.88 GB 1.38 GB 329.97 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 295.21 GB 1.38 GB 303.79 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 237.83 GB 1.38 GB 245.26 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Step 5 Preview (600B MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Step 5 Preview (600B MoE)?

Step 5 Preview (600B MoE) uses Step5ForConditionalGeneration with 92 layers (92-Layer Sparse MoE (352 Experts, Top-8 Active)), hidden_size=4096, 64 Q heads / 4 KV heads (head_dim=192), and vocab_size=128,896.

How much VRAM does Step 5 Preview (600B MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Step 5 Preview (600B MoE) requires 376.18 GB VRAM. Full FP16 requires 1234.84 GB.