Step-3.5-Flash (196B MoE) Architecture Parameters, Benchmarks & VRAM
196.8B total / 11B active MoE with 3-way Multi-Token Prediction (MTP-3) delivering up to 350 tok/s.
🧬 Step-3.5-Flash (196B MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (stepfun-ai/Step-3.5-Flash) config.json & Model Card
1. Backbone & Dimensions
step3p5- Total / Active Params
- 196.8B / 11B
- Hidden Size (d_model)
- 4,096
- FFN Intermediate Size
- 11,264
- MoE Expert Dim
- 1,280
- MoE Experts
- 288 Routed · Top-8 + 1 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=1e-6
- Native Tensor Dtype
- BF16 / FP8
2. Attention & Layer Layout
HYBRID- Num Hidden Layers
- 45 Layers (12 Global KV)
- Hidden Layer Layout
- 45-Layer 3:1 SWA-to-Full Hybrid Attention · 288 Routed Experts (Top-8 Active)
- Attention Heads (Q / KV)
- 64 Q / 8 KV (8:1 GQA)
- Head Dimension
- 128
- Linear / KV Compression
- 3-way Multi-Token Prediction (MTP-3) head boosts decode throughput to 100–300+ tok/s
- Vision Encoder / MTP
- MTP 3L
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 262,144 tokens
- RoPE Configuration
- YaRN Hybrid RoPE (原生 262,144 Context)
- Vocab Size
- 128,896 (Untied)
- Special Token IDs
- BOS: 0 · EOS: [1,2,128007]
📋 Step-3.5-Flash (196B MoE) Official API & System Card Specifications
Verified against official StepFun 阶跃星辰 developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- step-3.5-flash
- Release Date
- 2026-01-29
- 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
- 256,000 tokens (256K)
- Max Output Tokens
- 256,000 tokens
- Input Modalities
- Text
- Output Modalities
- Text
- Direct China Access
- ✅ Direct CN + RMB
- Deployment Mode
- Open-Weight + Cloud API
3. Official Pricing & Cache
CNY / 1M- Input / 1M Tokens
- ¥ 0.7
- Cached Input / 1M
- ¥ 0.14 -80%
- Output / 1M Tokens
- ¥ 2.1
- Tiered Pricing Notes
- StepFun: Official StepFun CN platform
🏆 Official Benchmark Results
Sourced from official System Card / Model Card & technical reports
International Mathematical Olympiad answer benchmark
LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.
A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.
Terminal-Bench 2.0 is an updated benchmark for testing AI agents' tool use ability to operate a computer via terminal.
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 Step-3.5-Flash (196B MoE) Full Benchmark Scorecard
7 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 2 items
A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.
Terminal-Bench 2.0 is an updated benchmark for testing AI agents' tool use ability to operate a computer via terminal.
🤖 Agents & Tools 2 items
τ-bench: A benchmark for tool-agent-user interaction in real-world domains. Tests language agents' ability to interact with users and follow domain-specific rules through dynamic conversations using API tools and policy guidelines across retail and airline domains.
BrowseComp is a benchmark comprising 1,266 questions that challenge AI agents to persistently navigate the internet in search of hard-to-find, entangled information.
📐 Math 2 items
All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.
IMO-AnswerBench is a benchmark for evaluating mathematical reasoning capabilities on International Mathematical Olympiad (IMO) problems, focusing on answer generation and verification.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ Step-3.5-Flash (196B MoE) Interactive VRAM & Token Speed Calculator
Physical 512 GB · Load 26.7%
MoE: reads only 6.67 GB active weights/step
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run step3.5:flashMac Studio M5 Ultra 512G Fits Step-3.5-Flash (196B MoE) · Cost & TCO
Approx. $9,499 (~$19/GB VRAM) · 190W TDP
📊 Step-3.5-Flash (196B MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 393.6 GB | 0.87 GB | 402.8 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 209.1 GB | 0.87 GB | 214.61 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 161.38 GB | 0.87 GB | 165.93 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 139.97 GB | 0.87 GB | 144.1 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 119.31 GB | 0.87 GB | 123.02 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 104.55 GB | 0.87 GB | 107.97 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 96.19 GB | 0.87 GB | 99.43 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 77.49 GB | 0.87 GB | 80.36 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for Step-3.5-Flash (196B MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of Step-3.5-Flash (196B MoE)?
Step-3.5-Flash (196B MoE) uses Step3p5ForCausalLM with 45 layers (45-Layer 3:1 SWA-to-Full Hybrid Attention · 288 Routed Experts (Top-8 Active)), hidden_size=4096, 64 Q heads / 8 KV heads (head_dim=128), and vocab_size=128,896.
How much VRAM does Step-3.5-Flash (196B MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), Step-3.5-Flash (196B MoE) requires 123.02 GB VRAM. Full FP16 requires 402.8 GB.