stepfun-ai/Step-3.5-Flash
838131k
1---2license: apache-2.03base_model:4- stepfun-ai/step-3.5-flash5library_name: transformers6---7 8# Step 3.5 Flash9 10<div align="center">11 12<div align="center" style="display: flex; justify-content: center; align-items: center;">13 <img src="stepfun.svg" width="25" style="margin-right: 10px;"/>14 <h1 style="margin: 0; border-bottom: none;">Step 3.5 Flash</h1>15</div>16 17[](https://github.com/stepfun-ai/Step-3.5-Flash)18[](https://huggingface.co/stepfun-ai/Step-3.5-Flash)19[](https://modelscope.cn/models/stepfun-ai/Step-3.5-Flash)20[](https://discord.gg/RcMJhNVAQc)21[](https://static.stepfun.com/blog/step-3.5-flash/)22[](https://arxiv.org/abs/2602.10604)23[]()24[](https://openrouter.ai/chat?models=stepfun/step-3.5-flash:free)25[](https://huggingface.co/spaces/stepfun-ai/Step-3.5-Flash)26</div>27 28 29 30## 1. Introduction 31 32**Step 3.5 Flash** ([visit website](https://static.stepfun.com/blog/step-3.5-flash/)) is our most capable open-source foundation model, engineered to deliver frontier reasoning and agentic capabilities with exceptional efficiency. Built on a sparse Mixture of Experts (MoE) architecture, it selectively activates only 11B of its 196B parameters per token. This "intelligence density" allows it to rival the reasoning depth of top-tier proprietary models, while maintaining the agility required for real-time interaction.33 34## 2. Key Capabilities35 36- **Deep Reasoning at Speed**: While chatbots are built for reading, agents must reason fast. Powered by 3-way Multi-Token Prediction (MTP-3), Step 3.5 Flash achieves a generation throughput of **100–300 tok/s** in typical usage (peaking at **350 tok/s** for single-stream coding tasks). This allows for complex, multi-step reasoning chains with immediate responsiveness.37 38- **A Robust Engine for Coding & Agents**: Step 3.5 Flash is purpose-built for agentic tasks, integrating a scalable RL framework that drives consistent self-improvement. It achieves **74.4% on SWE-bench Verified** and **51.0% on Terminal-Bench 2.0**, proving its ability to handle sophisticated, long-horizon tasks with unwavering stability.39 40- **Efficient Long Context**: The model supports a cost-efficient **256K context window** by employing a 3:1 Sliding Window Attention (SWA) ratio—integrating three SWA layers for every full-attention layer. This hybrid approach ensures consistent performance across massive datasets or long codebases while significantly reducing the computational overhead typical of standard long-context models.41 42- **Accessible Local Deployment**: Optimized for accessibility, Step 3.5 Flash brings elite-level intelligence to local environments. It runs securely on high-end consumer hardware (e.g., Mac Studio M4 Max, NVIDIA DGX Spark), ensuring data privacy without sacrificing performance.43 44## 3. Performance45 46Step 3.5 Flash delivers performance parity with leading closed-source systems while remaining open and efficient.47 4849 50Performance of Step 3.5 Flash measured across **Reasoning**, **Coding**, and **Agentic Abilities**. Open-source models (left) are sorted by their total parameter count, while top-tier proprietary models are shown on the right. xbench-DeepSearch scores are sourced from [official publications](https://xbench.org/agi/aisearch) for consistency. The shadowed bars represent the enhanced performance of Step 3.5 Flash using [Parallel Thinking](https://arxiv.org/pdf/2601.05593).51 52### Detailed Benchmarks53 54| Benchmark | Step 3.5 Flash | DeepSeek V3.2 | Kimi K2 Thinking / K2.5 | GLM-4.7 | MiniMax M2.1 | MiMo-V2 Flash |55| --- | --- | --- | --- | --- | --- | --- |56| # Activated Params | 11B | 37B | 32B | 32B | 10B | 15B |57| # Total Params (MoE) | 196B | 671B | 1T | 355B | 230B | 309B |58| Est. decoding cost @ 128K context, Hopper GPU** | **1.0x**<br>100 tok/s, MTP-3, EP8 | **6.0x**<br>33 tok/s, MTP-1, EP32 | **18.9x**<br>33 tok/s, no MTP, EP32 | **18.9x**<br>100 tok/s, MTP-3, EP8 | **3.9x**<br>100 tok/s, MTP-3, EP8 | **1.2x**<br>100 tok/s, MTP-3, EP8 |59| | | | **Agent** | | | |60| τ²-Bench | 88.2 | 80.3 (85.2*) | 74.3*/85.4* | 87.4 | 86.6* | 80.3 (84.1*) |61| BrowseComp | 51.6 | 51.4 | 41.5* / 60.6 | 52.0 | 47.4 | 45.4 |62| BrowseComp (w/ Context Manager) | 69.0 | 67.6 | 60.2/74.9 | 67.5 | 62.0 | 58.3 |63| BrowseComp-ZH | 66.9 | 65.0 | 62.3 / 62.3* | 66.6 | 47.8* | 51.2* |64| BrowseComp-ZH (w/ Context Manager) | 73.7 | — | —/— | — | — | — |65| GAIA (no file) | 84.5 | 75.1* | 75.6*/75.9* | 61.9* | 64.3* | 78.2* |66| xbench-DeepSearch (2025.05) | 83.7 | 78.0* | 76.0*/76.7* | 72.0* | 68.7* | 69.3* |67| xbench-DeepSearch (2025.10) | 56.3 | 55.7* | —/40+ | 52.3* | 43.0* | 44.0* |68| ResearchRubrics | 65.3 | 55.8* | 56.2*/59.5* | 62.0* | 60.2* | 54.3* |69| | | | **Reasoning** | | | |70| AIME 2025 | 97.3 | 93.1 | 94.5/96.1 | 95.7 | 83.0 | 94.1 (95.1*) |71| HMMT 2025 (Feb.) | 98.4 | 92.5 | 89.4/95.4 | 97.1 | 71.0* | 84.4 (95.4*) |72| HMMT 2025 (Nov.) | 94.0 | 90.2 | 89.2*/— | 93.5 | 74.3* | 91.0* |73| IMOAnswerBench | 85.4 | 78.3 | 78.6/81.8 | 82.0 | 60.4* | 80.9* |74| | | | **Coding** | | | |75| LiveCodeBench-V6 | 86.4 | 83.3 | 83.1/85.0 | 84.9 | — | 80.6 (81.6*) |76| SWE-bench Verified | 74.4 | 73.1 | 71.3/76.8 | 73.8 | 74.0 | 73.4 |77| Terminal-Bench 2.0 | 51.0 | 46.4 | 35.7*/50.8 | 41.0 | 47.9 | 38.5 |78 79**Notes**:801. "—" indicates the score is not publicly available or not tested.812. "*" indicates the original score was inaccessible or lower than our reproduced, so we report the evaluation under the same test conditions as Step 3.5 Flash to ensure fair comparability.823. **BrowseComp (with Context Manager)**: When the effective context length exceeds a predefined threshold, the agent resets the context and restarts the agent loop. By contrast, Kimi K2.5 and DeepSeek-V3.2 used a "discard-all" strategy.834. **Decoding Cost**: Estimates are based on a methodology similar to, but more accurate than, the approach described arxiv.org/abs/2507.1942784 85### Recommended Inference Parameters861. For general chat domain, we suggest: `temperature=0.6, top_p=0.95`872. For reasoning / agent scenario, we recommend: `temperature=1.0, top_p=0.95`.88 89## 4. Architecture Details90 91Step 3.5 Flash is built on a **Sparse Mixture-of-Experts (MoE)** transformer architecture, optimized for high throughput and low VRAM usage during inference.92 93### 4.1 Technical Specifications94 95| Component | Specification |96| :--- | :--- |97| **Backbone** | 45-layer Transformer (4,096 hidden dim) |98| **Context Window** | 256K |99| **Vocabulary** | 128,896 tokens |100| **Total Parameters** | **196.81B** (196B Backbone + 0.81B Head) |101| **Active Parameters** | **~11B** (per token generation) |102 103### 4.2 Mixture of Experts (MoE) Routing104 105Unlike traditional dense models, Step 3.5 Flash uses a fine-grained routing strategy to maximize efficiency:106- **Fine-Grained Experts**: 288 routed experts per layer + 1 shared expert (always active).107- **Sparse Activation**: Only the Top-8 experts are selected per token.108- **Result**: The model retains the "memory" of a 196B parameter model but executes with the speed of an 11B model. 109 110### 4.3 Multi-Token Prediction (MTP)111 112To improve inference speed, we utilize a specialized MTP Head consisting of a sliding-window attention mechanism and a dense Feed-Forward Network (FFN). This module predicts 4 tokens simultaneously in a single forward pass, significantly accelerating inference without degrading quality.113 114## 5. Quick Start115 116You can get started with Step 3.5 Flash in minutes using Cloud API via our supported providers.117 118### 5.1 Get Your API Key.119 120Sign up at [OpenRouter](https://openrouter.ai) or [platform.stepfun.ai](https://platform.stepfun.ai), and grab your API key. 121 122> OpenRouter now offers free trial for Step 3.5 Flash.123 124| Provider | Website | Base URL |125| :--- | :--- | :--- |126| OpenRouter | https://openrouter.ai | https://openrouter.ai/api/v1 |127| StepFun | https://platform.stepfun.ai | https://api.stepfun.ai/v1 |128 129### 5.2 Setup130 131Install the standard OpenAI SDK (compatible with both platforms).132 133```bash134pip install --upgrade "openai>=1.0"135```136 137Note: OpenRouter supports multiple SDKs. Learn more [here](https://openrouter.ai/docs/quickstart).138 139### 5.3 Implementation Example140 141This example shows starting a chat with Step 3.5 Flash.142 143```python144from openai import OpenAI145 146client = OpenAI(147 api_key="YOUR_API_KEY",148 base_url="https://api.stepfun.ai/v1", # or "https://openrouter.ai/api/v1"149 # Optional: OpenRouter headers for app rankings150 default_headers={151 "HTTP-Referer": "<YOUR_SITE_URL>", 152 "X-Title": "<YOUR_SITE_NAME>",153 }154)155 156completion = client.chat.completions.create(157 model="step-3.5-flash", # Use "stepfun/step-3.5-flash" for OpenRouter158 messages=[159 {160 "role": "system",161 "content": "You are an AI chat assistant provided by StepFun. You are good at Chinese, English, and many other languages.",162 },163 {164 "role": "user",165 "content": "Introduce StepFun's artificial intelligence capabilities."166 },167 ],168)169 170print(completion.choices[0].message.content)171```172 173## 6. Local Deployment174 175Step 3.5 Flash is optimized for local inference and supports industry-standard backends including vLLM, SGLang, Hugging Face Transformers and llama.cpp.176 177### 6.1 vLLM 178We recommend using the latest nightly build of vLLM.1791. Install vLLM.180 181```bash182# via Docker183docker pull vllm/vllm-openai:nightly184 185# or via pip (nightly wheels)186pip install -U vllm --pre \187 --index-url https://pypi.org/simple \188 --extra-index-url https://wheels.vllm.ai/nightly189```1902. Launch the server.191 192**Note**: Full MTP3 support is not yet available in vLLM. We are actively working on a Pull Request to integrate this feature, which is expected to significantly enhance decoding performance.193 194 - For fp8 model195```bash 196vllm serve <MODEL_PATH_OR_HF_ID> \197 --served-model-name step3p5-flash \198 --tensor-parallel-size 8 \199 --enable-expert-parallel \200 --disable-cascade-attn \201 --reasoning-parser step3p5 \202 --enable-auto-tool-choice \203 --tool-call-parser step3p5 \204 --hf-overrides '{"num_nextn_predict_layers": 1}' \205 --speculative_config '{"method": "step3p5_mtp", "num_speculative_tokens": 1}' \206 --trust-remote-code \207 --quantization fp8208```209 210 - For bf16 model211```bash212vllm serve <MODEL_PATH_OR_HF_ID> \213 --served-model-name step3p5-flash \214 --tensor-parallel-size 8 \215 --enable-expert-parallel \216 --disable-cascade-attn \217 --reasoning-parser step3p5 \218 --enable-auto-tool-choice \219 --tool-call-parser step3p5 \220 --hf-overrides '{"num_nextn_predict_layers": 1}' \221 --speculative_config '{"method": "step3p5_mtp", "num_speculative_tokens": 1}' \222 --trust-remote-code 223```224You can also refer to the [Step-3.5-Flash](https://github.com/vllm-project/recipes/blob/main/StepFun/Step-3.5-Flash.md) recipe.225 226### 6.2 SGLang227 2281. Install SGLang.229```bash230# via Docker231docker pull lmsysorg/sglang:dev-pr-18084232# or from source (pip)233pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git"234```235 2362. Launch the server.237 - For bf16 model238 239```bash240sglang serve --model-path <MODEL_PATH_OR_HF_ID> \241 --served-model-name step3p5-flash \242 --tp-size 8 \243 --tool-call-parser step3p5 \244 --reasoning-parser step3p5 \245 --speculative-algorithm EAGLE \246 --speculative-num-steps 3 \247 --speculative-eagle-topk 1 \248 --speculative-num-draft-tokens 4 \249 --enable-multi-layer-eagle \250 --host 0.0.0.0 \251 --port 8000252```253 - For fp8 model254```bash255sglang serve --model-path <MODEL_PATH_OR_HF_ID> \256 --served-model-name step3p5-flash \257 --tp-size 8 \258 --ep-size 8 \259 --tool-call-parser step3p5 \260 --reasoning-parser step3p5 \261 --speculative-algorithm EAGLE \262 --speculative-num-steps 3 \263 --speculative-eagle-topk 1 \264 --speculative-num-draft-tokens 4 \265 --enable-multi-layer-eagle \266 --host 0.0.0.0 \267 --port 8000268```269 270### 6.3 Transformers (Debug / Verification)271 272Use this snippet for quick functional verification. For high-throughput serving, use vLLM or SGLang.273```python274from transformers import AutoModelForCausalLM, AutoTokenizer275 276MODEL_PATH = "<MODEL_PATH_OR_HF_ID>"277 278# 1. Setup279tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)280model = AutoModelForCausalLM.from_pretrained(281 MODEL_PATH,282 trust_remote_code=True,283 torch_dtype="auto",284 device_map="auto",285)286 287# 2. Prepare Input288messages = [{"role": "user", "content": "Explain the significance of the number 42."}]289inputs = tokenizer.apply_chat_template(290 messages,291 tokenize=True,292 add_generation_prompt=True,293 return_dict=True,294 return_tensors="pt",295).to(model.device)296 297# 3. Generate298generated_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)299output_text = tokenizer.decode(generated_ids[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)300 301print(output_text)302```303 304### 6.4 llama.cpp305 306#### System Requirements307- GGUF Model Weights(int4): 111.5 GB308- Runtime Overhead: ~7 GB309- Minimum VRAM: 120 GB (e.g., Mac studio, DGX-Spark, AMD Ryzen AI Max+ 395)310- Recommended: 128GB unified memory311#### Steps3121. Use official llama.cpp:313> the folder `Step-3.5-Flash/tree/main/llama.cpp` is **obsolete**314```bash315git clone https://github.com/ggml-org/llama.cpp316cd llama.cpp317```3182. Build llama.cpp on Mac:319```bash320cmake -S . -B build-macos \321 -DCMAKE_BUILD_TYPE=Release \322 -DGGML_METAL=ON \323 -DGGML_ACCELERATE=ON \324 -DLLAMA_BUILD_EXAMPLES=ON \325 -DLLAMA_BUILD_COMMON=ON \326 -DGGML_LTO=ON327cmake --build build-macos -j8328```3293. Build llama.cpp on DGX-Spark:330```bash331cmake -S . -B build-cuda \332 -DCMAKE_BUILD_TYPE=Release \333 -DGGML_CUDA=ON \334 -DGGML_CUDA_GRAPHS=ON \335 -DLLAMA_CURL=OFF \336 -DLLAMA_BUILD_EXAMPLES=ON \337 -DLLAMA_BUILD_COMMON=ON338cmake --build build-cuda -j8339```3404. Build llama.cpp on AMD Windows:341```bash342cmake -S . -B build-vulkan \343 -DCMAKE_BUILD_TYPE=Release \344 -DLLAMA_CURL=OFF \345 -DGGML_OPENMP=ON \346 -DGGML_VULKAN=ON347cmake --build build-vulkan -j8348```3495. Run with llama-cli350```bash351./llama-cli -m step3.5_flash_Q4_K_S.gguf -c 16384 -b 2048 -ub 2048 -fa on --temp 1.0 -p "What's your name?"352```3536. Test performance with llama-batched-bench:354```bash355./llama-batched-bench -m step3.5_flash_Q4_K_S.gguf -c 32768 -b 2048 -ub 2048 -npp 0,2048,8192,16384,32768 -ntg 128 -npl 1356```357 358## 7. Using Step 3.5 Flash on Agent Platforms359 360### 7.1 Claude Code & Codex361It's straightforward to add Step 3.5 Flash to the list of models in most coding environments. See below for the instructions for configuring Claude Code and Codex to use Step 3.5 Flash.362 363#### 7.1.1 Prerequisites364Sign up at StepFun.ai or OpenRouter and grab an API key, as mentioned in the Quick Start.365 366#### 7.1.2 Environment setup367Claude Code and Codex rely on Node.js. We recommend installing Node.js version > v20. You can install Node via nvm.368 369**Mac/Linux**:370```bash371# Install nvm on Mac/Linux via curl：372# Step 1373curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.39.0/install.sh | bash374 375# Copy the full command376export NVM_DIR="$HOME/.nvm"377[ -s "$NVM_DIR/nvm.sh" ] && \. "$NVM_DIR/nvm.sh" # This loads nvm378[ -s "$NVM_DIR/bash_completion" ] && \. "$NVM_DIR/bash_completion"379 380# Users in China can set up npm mirror381config set registry https://registry.npmmirror.com382 383# Step 2384nvm install v22385 386# Make sure Node.js is installed387node --version388 389npm --version390```391 392**Windows**:393You can download the installation file (`nvm-setup.exe`) from [https://github.com/coreybutler/nvm-windows/releases](https://github.com/coreybutler/nvm-windows/releases). Follow the instructions to install nvm. Run nvm commands to make sure it is installed.394 395#### 7.1.3 Use Step 3.5 Flash on Claude Code396 3971. Install Claude Code.398```bash399# install claude code via npm400npm install -g @anthropic-ai/claude-code401 402# test if the installation is successful403claude --version 404```405 4062. Configure Claude Code.407 408 To accommodate diverse workflows in Claude Code, we support both **Anthropic-style** and **OpenAI-style** APIs. 409 410 **Option A: Anthropic API style**:411 412 > If you intend to use the **OpenRouter** API, refer to the OpenRouter integration guide. 413 414 Step 1: Edit Claude Settings. Update `~/.claude/settings.json`.415 > You only need to modify the fields shown below. Leave the rest of the file unchanged.416 417 ```json418 {419 "env": {420 "ANTHROPIC_API_KEY": "API_KEY_from_StepFun",421 "ANTHROPIC_BASE_URL": "https://api.stepfun.ai/"422 },423 "model": "step-3.5-flash"424 }425 ```426 Step 2: Start Claude Code.427 428 Save the file, and then start Claude Code. Run `/status` to confirm the model and base URL.429 430 ```txt431 ❯ /status432 ─────────────────────────────────────────────────────────────────────────────────433 Settings: Status Config Usage (←/→ or tab to cycle)434 435 Version: 2.1.1436 Session name: /rename to add a name437 Session ID: 676dae61-259d-4eef-8c2f-0f1641600553438 cwd: /Users/step-test/439 Auth token: none440 API key: ANTHROPIC_API_KEY441 Anthropic base URL: https://api.stepfun.ai/442 443 Model: step-3.5-flash444 Setting sources: User settings445 ```446 447 **Option B: OpenAI API style**448 449 > Note: OpenAI API style here refers to the `chat/completions/` format.450 451 > We recommend using `claude-code-router`. For details, see [https://github.com/musistudio/claude-code-router](https://github.com/musistudio/claude-code-router).452 453 After Claude Code is installed, install `claude-code-router` :454 455 ```bash456 # install ccr via npm457 npm install -g @musistudio/claude-code-router458 459 # validate it is installed460 ccr -v461 ```462 463 Add the following configurations to `~/.claude-code-router/config.json`.464 465 ```json466 {467 "PORT": 3456,468 "Providers": [469 {470 "name": "stepfun-api",471 "api_base_url": "https://api.stepfun.com/v1/chat/completions",472 "api_key": "StepFun_API_KEY",473 "models": ["step-3.5-flash"],474 "transformer":{475 "step-3.5-flash": { "use": ["OpenAI"]}476 }477 }478 ],479 "Router": {480 "default": "stepfun-api,step-3.5-flash",481 "background": "stepfun-api,step-3.5-flash",482 "think": "stepfun-api,step-3.5-flash",483 "longContext": "stepfun-api,step-3.5-flash",484 "webSearch": "stepfun-api,step-3.5-flash"485 }486 }487 ```488 You can now start Claude Code:489 490 ```bash491 # Start Claude492 ccr code 493 494 # restart ccr if configs are changed495 ccr restart 496 ```497 498#### 7.1.4 Use Step 3.5 Flash on Codex4991. Install Codex500```bash501# Install codex via npm502npm install -g @openai/codex503 504# Test if it is installed505codex --version506```507 5082. Configure Codex509Add the following settings to `~/.codex/config.toml`, keeping the rest of the settings as they are.510 511```json512model="step-3.5-flash"513model_provider = "stepfun-chat"514preferred_auth_method = "apikey"515 516# configure the provider517[model_providers.stepfun-chat]518name = "OpenAI using response"519base_url = "https://api.stepfun.com/v1"520env_key = "OPENAI_API_KEY"521wire_api = "chat"522query_params = {}523```524 525For Codex, `wire_api` only supports `chat` . If you use the `responses` mode, you'll need to change to `chat`. Please also switch `model_provider` to the newly configured `stepfun-chat`.526 527When finishing the configuration, run codex in a new Terminal window to start Codex. Run `/status` to check the configuration.528 529```bash530/status531📂 Workspace532 • Path: /Users/step-test/533 • Approval Mode: on-request534 • Sandbox: workspace-write535 • AGENTS files: (none)536 537🧠 Model538 • Name: step-3.5-flash539 • Provider: Stepfun-chat540 541💻 Client542 • CLI Version: 0.40.0543```544 545#### 7.1.5 Use Step 3.5 Flash on Step-DeepResearch (DeepResearch)5461. Use the reference environment setup below and configure `MODEL_NAME` to `Step-3.5-Flash`. [https://github.com/stepfun-ai/StepDeepResearch?tab=readme-ov-file#1-environment-setup](https://github.com/stepfun-ai/StepDeepResearch?tab=readme-ov-file#1-environment-setup)547 548 549## 8. Known Issues and Future Directions550 5511. **Token Efficiency**. Step 3.5 Flash achieves frontier-level agentic intelligence but currently relies on longer generation trajectories than Gemini 3.0 Pro to reach comparable quality.5522. **Efficient Universal Mastery**. We aim to unify generalist versatility with deep domain expertise. To achieve this efficiently, we are advancing variants of on-policy distillation, allowing the model to internalize expert behaviors with higher sample efficiency.5533. **RL for More Agentic Tasks**. While Step 3.5 Flash demonstrates competitive performance on academic agentic benchmarks, the next frontier of agentic AI necessitates the application of RL to intricate, expert-level tasks found in professional work, engineering, and research.5544. **Operational Scope and Constraints**. Step 3.5 Flash is tailored for coding and work-centric tasks, but may experience reduced stability during distribution shifts. This typically occurs in highly specialized domains or long-horizon, multi-turn dialogues, where the model may exhibit repetitive reasoning, mixed-language outputs, or inconsistencies in time and identity awareness.555 556## 9. Co-Developing the Future557 558We view our roadmap as a living document, evolving continuously based on real-world usage and developer feedback.559As we work to shape the future of AGI by expanding broad model capabilities, we want to ensure we are solving the right problems. We invite you to be part of this continuous feedback loop—your insights directly influence our priorities.560 561- **Join the Conversation**: Our Discord community is the primary hub for brainstorming future architectures, proposing capabilities, and getting early access updates 🚀562- **Report Friction**: Encountering limitations? You can open an issue on GitHub or flag it directly in our Discord support channels.563 564## 📜 Citation565 566If you find this project useful in your research, please cite our technical report:567 568```tex569@misc{huang2026step35flashopen,570 title={Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters}, 571 author={Ailin Huang and Ang Li and Aobo Kong and Bin Wang and Binxing Jiao and Bo Dong and Bojun Wang and Boyu Chen and Brian Li and Buyun Ma and Chang Su and Changxin Miao and Changyi Wan and Chao Lou and Chen Hu and Chen Xu and Chenfeng Yu and Chengting Feng and Chengyuan Yao and Chunrui Han and Dan Ma and Dapeng Shi and Daxin Jiang and Dehua Ma and Deshan Sun and Di Qi and Enle Liu and Fajie Zhang and Fanqi Wan and Guanzhe Huang and Gulin Yan and Guoliang Cao and Guopeng Li and Han Cheng and Hangyu Guo and Hanshan Zhang and Hao Nie and Haonan Jia and Haoran Lv and Hebin Zhou and Hekun Lv and Heng Wang and Heung-Yeung Shum and Hongbo Huang and Hongbo Peng and Hongyu Zhou and Hongyuan Wang and Houyong Chen and Huangxi Zhu and Huimin Wu and Huiyong Guo and Jia Wang and Jian Zhou and Jianjian Sun and Jiaoren Wu and Jiaran Zhang and Jiashu Lv and Jiashuo Liu and Jiayi Fu and Jiayu Liu and Jie Cheng and Jie Luo and Jie Yang and Jie Zhou and Jieyi Hou and Jing Bai and Jingcheng Hu and Jingjing Xie and Jingwei Wu and Jingyang Zhang and Jishi Zhou and Junfeng Liu and Junzhe Lin and Ka Man Lo and Kai Liang and Kaibo Liu and Kaijun Tan and Kaiwen Yan and Kaixiang Li and Kang An and Kangheng Lin and Lei Yang and Liang Lv and Liang Zhao and Liangyu Chen and Lieyu Shi and Liguo Tan and Lin Lin and Lina Chen and Luck Ma and Mengqiang Ren and Michael Li and Ming Li and Mingliang Li and Mingming Zhang and Mingrui Chen and Mitt Huang and Na Wang and Peng Liu and Qi Han and Qian Zhao and Qinglin He and Qinxin Du and Qiuping Wu and Quan Sun and Rongqiu Yang and Ruihang Miao and Ruixin Han and Ruosi Wan and Ruyan Guo and Shan Wang and Shaoliang Pang and Shaowen Yang and Shengjie Fan and Shijie Shang and Shiliang Yang and Shiwei Li and Shuangshuang Tian and Siqi Liu and Siye Wu and Siyu Chen and Song Yuan and Tiancheng Cao and Tianchi Yue and Tianhao Cheng and Tianning Li and Tingdan Luo and Wang You and Wei Ji and Wei Yuan and Wei Zhang and Weibo Wu and Weihao Xie and Wen Sun and Wenjin Deng and Wenzhen Zheng and Wuxun Xie and Xiangfeng Wang and Xiangwen Kong and Xiangyu Liu and Xiangyu Zhang and Xiaobo Yang and Xiaojia Liu and Xiaolan Yuan and Xiaoran Jiao and Xiaoxiao Ren and Xiaoyun Zhang and Xin Li and Xin Liu and Xin Wu and Xing Chen and Xingping Yang and Xinran Wang and Xu Zhao and Xuan He and Xuanti Feng and Xuedan Cai and Xuqiang Zhou and Yanbo Yu and Yang Li and Yang Xu and Yanlin Lai and Yanming Xu and Yaoyu Wang and Yeqing Shen and Yibo Zhu and Yichen Lv and Yicheng Cao and Yifeng Gong and Yijing Yang and Yikun Yang and Yin Zhao and Yingxiu Zhao and Yinmin Zhang and Yitong Zhang and Yixuan Zhang and Yiyang Chen and Yongchi Zhao and Yongshen Long and Yongyao Wang and Yousong Guan and Yu Zhou and Yuang Peng and Yuanhao Ding and Yuantao Fan and Yuanzhen Yang and Yuchu Luo and Yudi Zhao and Yue Peng and Yueqiang Lin and Yufan Lu and Yuling Zhao and Yunzhou Ju and Yurong Zhang and Yusheng Li and Yuxiang Yang and Yuyang Chen and Yuzhu Cai and Zejia Weng and Zetao Hong and Zexi Li and Zhe Xie and Zheng Ge and Zheng Gong and Zheng Zeng and Zhenyi Lu and Zhewei Huang and Zhichao Chang and Zhiguo Huang and Zhiheng Hu and Zidong Yang and Zili Wang and Ziqi Ren and Zixin Zhang and Zixuan Wang},572 year={2026},573 eprint={2602.10604},574 archivePrefix={arXiv},575 primaryClass={cs.CL},576 url={https://arxiv.org/abs/2602.10604}, 577}578```579 580## License581This project is open-sourced under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).