[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-75130":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":9,"rankLanguage":9,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":22,"hasPages":22,"topics":24,"createdAt":9,"pushedAt":9,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":15,"starSnapshotCount":15,"syncStatus":28,"lastSyncTime":29,"discoverSource":30},75130,"Step-3.5-Flash","stepfun-ai\u002FStep-3.5-Flash","stepfun-ai","Fast, Sharp & Reliable Agentic Intelligence",null,"C++",2076,84,7,15,0,3,16,48,9,76.59,"Apache License 2.0",false,"main",[],"2026-06-12 04:01:17","\u003C!-- ---\nlicense: apache-2.0\nbase_model:\n  - stepfun-ai\u002Fstep-3.5-flash\n--- -->\n\n\u003Cdiv align=\"center\"> \n  \u003Ch1 style=\"margin: 0; border-bottom: none;\"> \u003Cimg src=\"assets\u002Fstepfun-1.png\" width=\"25\" style=\"margin-right: 10px;\"\u002F>  Step 3.5 Flash\u003C\u002Fh1>\n\u003C\u002Fdiv>\n\n\u003Cp align=\"center\">\n    \u003Cstrong>English\u003C\u002Fstrong>  |  \u003Ca href=\"README.zh-CN.md\">简体中文\u003C\u002Fa>\n\u003C\u002Fp>\n\n\n\u003Cp align=\"center\">\n    \u003Ca href=\".\u002Fcookbooks\u002Fopenclaw\">OpenClaw Guide\u003C\u002Fa>  |  \u003Ca href=\".\u002Fcookbooks\u002Fclaude-code-best-practices\u002FREADME.en.md\">Claude Code Guide\u003C\u002Fa>  |  \u003Ca href=\".\u002Fcookbooks\u002Froo-code-integration-guide\">Roo Code Guide\u003C\u002Fa>  |  \u003Ca href=\".\u002Fcookbooks\u002Fhybrid-local-agent-macos\">Local Agent Guide\u003C\u002Fa>\n\u003C\u002Fp>\n\n\n\n[![Hugging Face](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20HF-StepFun\u002FSTEP3p5-preview)](https:\u002F\u002Fhuggingface.co\u002Fstepfun-ai\u002FStep-3.5-Flash\u002Ftree\u002Fmain)\n[![ModelScope](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FModelScope-StepFun\u002FSTEP3p5-preview)](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002Fstepfun-ai\u002FStep-3.5-Flash)\n[![Discord](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDiscord-Join-5865F2?logo=discord&logoColor=white)](https:\u002F\u002Fdiscord.gg\u002FRcMJhNVAQc)\n[![Paper](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-Arxiv-red)](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2602.10604)\n[![Webpage](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWebpage-Blog-blue)](https:\u002F\u002Fstatic.stepfun.com\u002Fblog\u002Fstep-3.5-flash\u002F)\n[![License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache%202.0-green)]()\n[![Chat with the model on OpenRouter](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FChat%20with%20the%20model-OpenRouter-5B3DF5?logo=chatbot&logoColor=white)](https:\u002F\u002Fopenrouter.ai\u002Fchat?models=stepfun\u002Fstep-3.5-flash:free)\n[![Chat with the model on HuggingfaceSpace](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FChat%20with%20the%20model-HuggingfaceSpace-5B3DF5?logo=chatbot&logoColor=white)](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fstepfun-ai\u002FStep-3.5-Flash)\n[![OpenClaw Integration](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FAgent-OpenClaw-purple?logo=robot&logoColor=white)](.\u002Fcookbooks\u002Fopenclaw)\n\n\u003C\u002Fdiv>\n\n## 1. Introduction \n\n**Step 3.5 Flash** ([visit website](https:\u002F\u002Fstatic.stepfun.com\u002Fblog\u002Fstep-3.5-flash\u002F)) 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.\n\n### Contents\n- [Key Capabilities](#2-key-capabilities)\n- [Performance](#3-performance)\n- [Architecture Details](#4-architecture-details)\n- [Quick Start](#5-quick-start)\n- [Local Deployment](#6-local-deployment)\n- [Using Step 3.5 Flash on Agent Platforms](#7-using-step-35-flash-on-agent-platforms)\n- [Cookbooks](#8-cookbooks)\n- [Known Issues and Future Directions](#9-known-issues-and-future-directions)\n- [Co-Developing the Future](#10-co-developing-the-future)\n- [License](#license)\n\n## 2. Key Capabilities\n\n- **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\u002Fs** in typical usage (peaking at **350 tok\u002Fs** for single-stream coding tasks). This allows for complex, multi-step reasoning chains with immediate responsiveness.\n\n- **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.\n\n- **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.\n\n- **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.\n\n## 3. Performance\n\nStep 3.5 Flash delivers performance parity with leading closed-source systems while remaining open and efficient.\n\n![](assets\u002Fstep-bar-chart.png)\n\nPerformance of Step 3.5 Flash measured across **Reasoning**, **Coding**, and **Agentic Capabilities**. 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:\u002F\u002Fxbench.org\u002Fagi\u002Faisearch) for consistency. The shadowed bars represent the enhanced performance of Step 3.5 Flash using [Parallel Thinking](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2601.05593).\n\n### Detailed Benchmarks\n\n| Benchmark | Step 3.5 Flash | DeepSeek V3.2 | Kimi K2 Thinking \u002F K2.5 | GLM-4.7 | MiniMax M2.1 | MiMo-V2 Flash |\n| --- | --- | --- | --- | --- | --- | --- |\n| # Activated Params | 11B | 37B | 32B | 32B | 10B | 15B |\n| # Total Params (MoE) | 196B | 671B | 1T | 355B | 230B | 309B |\n| Est. decoding cost @ 128K context, Hopper GPU** | **1.0x**\u003Cbr>100 tok\u002Fs, MTP-3, EP8 | **6.0x**\u003Cbr>33 tok\u002Fs, MTP-1, EP32 | **18.9x**\u003Cbr>33 tok\u002Fs, no MTP, EP32 | **18.9x**\u003Cbr>100 tok\u002Fs, MTP-3, EP8 | **3.9x**\u003Cbr>100 tok\u002Fs, MTP-3, EP8 | **1.2x**\u003Cbr>100 tok\u002Fs, MTP-3, EP8 |\n| | | | **Agent** | | | |\n| τ²-Bench | 88.2 | 80.3 (85.2*) | 74.3*\u002F85.4* | 87.4 | 86.6* | 80.3 (84.1*) |\n| BrowseComp | 51.6 | 51.4 | 41.5* \u002F 60.6 | 52.0 | 47.4 | 45.4 |\n| BrowseComp (w\u002F Context Manager) | 69.0 | 67.6 | 60.2\u002F74.9 | 67.5 | 62.0 | 58.3 |\n| BrowseComp-ZH | 66.9 | 65.0 | 62.3 \u002F 62.3* | 66.6 | 47.8* | 51.2* |\n| BrowseComp-ZH (w\u002F Context Manager) | 73.7 | — | —\u002F— | — | — | — |\n| GAIA (no file) | 84.5 | 75.1* | 75.6*\u002F75.9* | 61.9* | 64.3* | 78.2* |\n| xbench-DeepSearch (2025.05) | 83.7 | 78.0* | 76.0*\u002F76.7* | 72.0* | 68.7* | 69.3* |\n| xbench-DeepSearch (2025.10) | 56.3 | 55.7* | —\u002F40+ | 52.3* | 43.0* | 44.0* |\n| ResearchRubrics | 65.3 | 55.8* | 56.2*\u002F59.5* | 62.0* | 60.2* | 54.3* |\n| | | | **Reasoning** | | | |\n| AIME 2025 | 97.3 | 93.1 | 94.5\u002F96.1 | 95.7 | 83.0 | 94.1 (95.1*) |\n| HMMT 2025 (Feb.) | 98.4 | 92.5 | 89.4\u002F95.4 | 97.1 | 71.0* | 84.4 (95.4*) |\n| HMMT 2025 (Nov.) | 94.0 | 90.2 | 89.2*\u002F— | 93.5 | 74.3* | 91.0* |\n| IMOAnswerBench | 85.4 | 78.3 | 78.6\u002F81.8 | 82.0 | 60.4* | 80.9* |\n| | | | **Coding** | | | |\n| LiveCodeBench-V6 | 86.4 | 83.3 | 83.1\u002F85.0 | 84.9 | — | 80.6 (81.6*) |\n| SWE-bench Verified | 74.4 | 73.1 | 71.3\u002F76.8 | 73.8 | 74.0 | 73.4 |\n| Terminal-Bench 2.0 | 51.0 | 46.4 | 35.7*\u002F50.8 | 41.0 | 47.9 | 38.5 |\n\n## Notes\n\n- \"—\" indicates the score is not publicly available or not tested.\n- \"*\" 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.\n- 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.)\n- In decoding cost section, decoding is estimated using a similar but more accurate approach than [arxiv.org\u002Fabs\u002F2507.19427](https:\u002F\u002Farxiv.org\u002Fabs\u002F2507.19427).\n\n## 4. Architecture Details\n\nStep 3.5 Flash is built on a **Sparse Mixture-of-Experts (MoE)** transformer architecture, optimized for high throughput and low VRAM usage during inference.\n\n### 4.1 Technical Specifications\n\n| Component | Specification |\n| :--- | :--- |\n| **Backbone** | 45-layer Transformer (4,096 hidden dim) |\n| **Context Window** | 256K |\n| **Vocabulary** | 128,896 tokens |\n| **Total Parameters** | **196.81B** (196B Backbone + 0.81B Head) |\n| **Active Parameters** | **~11B** (per token generation) |\n\n### 4.2 Mixture of Experts (MoE) Routing\n\nUnlike traditional dense models, Step 3.5 Flash uses a fine-grained routing strategy to maximize efficiency:\n- **Fine-Grained Experts**: 288 routed experts per layer + 1 shared expert (always active).\n- **Sparse Activation**: Only the Top-8 experts are selected per token.\n- **Result**: The model retains the \"memory\" of a 196B parameter model but executes with the speed of an 11B model. \n\n### 4.3 Multi-Token Prediction (MTP)\n\nTo 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.\n\n## 5. Quick Start\n\nYou can get started with Step 3.5 Flash in minutes using Cloud API via our supported providers.\n\n### 5.1 OpenRouter\n\n[OpenRouter](https:\u002F\u002Fopenrouter.ai) provides uniform access to Step 3.5 Flash with both free and paid tiers.\n\n**Models:**\n- **Free Tier:** [`stepfun\u002Fstep-3.5-flash:free`](https:\u002F\u002Fopenrouter.ai\u002Fstepfun\u002Fstep-3.5-flash:free)\n- **Standard Tier:** [`stepfun\u002Fstep-3.5-flash`](https:\u002F\u002Fopenrouter.ai\u002Fstepfun\u002Fstep-3.5-flash)\n\n**Configuration:**\n- **Base URL:** `https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1`\n- **API Key:** Sign up at [OpenRouter](https:\u002F\u002Fopenrouter.ai) to get your key.\n\n### 5.2 StepFun Platform\n\nStepFun offers official API endpoints for both International and Chinese users.\n\n| Region | Website | Base URL |\n| :--- | :--- | :--- |\n| **International** | [platform.stepfun.ai](https:\u002F\u002Fplatform.stepfun.ai) | `https:\u002F\u002Fapi.stepfun.ai\u002Fv1` |\n| **China** | [platform.stepfun.com](https:\u002F\u002Fplatform.stepfun.com) | `https:\u002F\u002Fapi.stepfun.com\u002Fv1` |\n\n> **Note:** China platform requires +86 phone number verification.\n\n### 5.3 OpenClaw (Recommended)\n\nFor a full agentic experience, we recommend using [OpenClaw](https:\u002F\u002Fopenclaw.ai).\n👉 **[View the OpenClaw Cookbook](.\u002Fcookbooks\u002Fopenclaw)** to get started in minutes.\n\n### 5.4 Implementation Example\n\nInstall the standard OpenAI SDK (compatible with both platforms):\n\n```bash\npip install --upgrade \"openai>=1.0\"\n```\n\n**Python Example:**\n\n```python\nfrom openai import OpenAI\n\n# Configuration for OpenRouter\n# base_url = \"https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1\"\n# api_key = \"sk-or-...\" \n# model = \"stepfun\u002Fstep-3.5-flash\"\n\n# Configuration for StepFun (International)\nbase_url = \"https:\u002F\u002Fapi.stepfun.ai\u002Fv1\" \n# For China: base_url = \"https:\u002F\u002Fapi.stepfun.com\u002Fv1\"\napi_key = \"your-stepfun-api-key\"\nmodel = \"step-3.5-flash\"\n\nclient = OpenAI(api_key=api_key, base_url=base_url)\n\ncompletion = client.chat.completions.create(\n    model=model,\n    messages=[\n        {\"role\": \"system\", \"content\": \"You are a helpful assistant.\"},\n        {\"role\": \"user\", \"content\": \"Hello! Introduce yourself.\"}\n    ]\n)\n\nprint(completion.choices[0].message.content)\n```\n\n## 6. Local Deployment\n\nStep 3.5 Flash is optimized for local inference and supports industry-standard backends including vLLM, SGLang, Hugging Face Transformers and llama.cpp.\n\n### 6.1 vLLM \nWe recommend using the latest nightly build of vLLM.\n1. Install vLLM.\n\n```bash\n# via Docker\ndocker pull vllm\u002Fvllm-openai:nightly\n\n# or via pip (nightly wheels)\npip install -U vllm --pre \\\n  --index-url https:\u002F\u002Fpypi.org\u002Fsimple \\\n  --extra-index-url https:\u002F\u002Fwheels.vllm.ai\u002Fnightly\n```\n\n**Note**: Bug fixes for tool parser and reasoning parser, as well as support for the `v1\u002Fmessages` interface, are being merged into vLLM. In the meantime, you can use the `vllm\u002Fvllm-openai:v0.15.1-x86_64` image with `step3.5_vllm_v0.15.1.patch` for deployment. The patch file `step3.5_vllm_v0.15.1.patch` is provided in this repository; download or copy it and ensure it is accessible from the directory where you run `git apply` (or use its absolute path).\n```bash \n# via Docker\n# refer to \"step3.5_vllm_v0.15.1.Dockerfile\"\n\n# or via pip \npip install -U vllm==0.15.1\ncd \u002Fpath\u002Fto\u002Flib\u002Fpython3.12\u002Fsite-packages  # Replace with the parent dir of vLLM installation path.\ngit apply \u002Fpath\u002Fto\u002Fstep3.5_vllm_v0.15.1.patch  # Replace with the actual path to the patch file (e.g., from the repository root).\n```\n\n2. Launch the server.\n\n**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.\n\n  - For fp8 model\n```bash  \nvllm serve \u003CMODEL_PATH_OR_HF_ID> \\\n  --served-model-name step3p5-flash \\\n  --tensor-parallel-size 8 \\\n  --enable-expert-parallel \\\n  --disable-cascade-attn \\\n  --reasoning-parser step3p5 \\\n  --enable-auto-tool-choice \\\n  --tool-call-parser step3p5 \\\n  --hf-overrides '{\"num_nextn_predict_layers\": 1}' \\\n  --speculative_config '{\"method\": \"step3p5_mtp\", \"num_speculative_tokens\": 1}' \\\n  --trust-remote-code \\\n  --quantization fp8\n```\n\n  - For bf16 model\n```bash\nvllm serve \u003CMODEL_PATH_OR_HF_ID> \\\n  --served-model-name step3p5-flash \\\n  --tensor-parallel-size 8 \\\n  --enable-expert-parallel \\\n  --disable-cascade-attn \\\n  --reasoning-parser step3p5 \\\n  --enable-auto-tool-choice \\\n  --tool-call-parser step3p5 \\\n  --hf-overrides '{\"num_nextn_predict_layers\": 1}' \\\n  --speculative_config '{\"method\": \"step3p5_mtp\", \"num_speculative_tokens\": 1}' \\\n  --trust-remote-code    \n```\n\n### 6.2 SGLang\n\n1. Install SGLang.\n```bash\n# via Docker\ndocker pull lmsysorg\u002Fsglang:dev-pr-18084\n# or from source (pip)\npip install \"sglang[all] @ git+https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang.git\"\n```\n\n2. Launch the server.\n  - For bf16 model\n```\nSGLANG_ENABLE_SPEC_V2=1\nsglang serve \\\n  --model-path \u003CMODEL_PATH_OR_HF_ID> \\\n  --served-model-name step3p5-flash \\\n  --tp-size 8 \\\n  --tool-call-parser step3p5 \\\n  --reasoning-parser step3p5 \\\n  --speculative-algorithm EAGLE \\\n  --speculative-num-steps 3 \\\n  --speculative-eagle-topk 1 \\\n  --speculative-num-draft-tokens 4 \\\n  --enable-multi-layer-eagle \\\n  --host 0.0.0.0 \\\n  --port 8000\n```\n  - For fp8 model\n```bash\nSGLANG_ENABLE_SPEC_V2=1\nsglang serve \\\n  --model-path \u003CMODEL_PATH_OR_HF_ID> \\\n  --served-model-name step3p5-flash \\\n  --tp-size 8 \\\n  --ep-size 8 \\\n  --tool-call-parser step3p5 \\\n  --reasoning-parser step3p5 \\\n  --speculative-algorithm EAGLE \\\n  --speculative-num-steps 3 \\\n  --speculative-eagle-topk 1 \\\n  --speculative-num-draft-tokens 4 \\\n  --enable-multi-layer-eagle \\\n  --host 0.0.0.0 \\\n  --port 8000\n```\n\n### 6.3 Transformers (Debug \u002F Verification)\n\nUse this snippet for quick functional verification. For high-throughput serving, use vLLM or SGLang.\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nMODEL_PATH = \"\u003CMODEL_PATH_OR_HF_ID>\"\n\n# 1. Setup\ntokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)\nmodel = AutoModelForCausalLM.from_pretrained(\n    MODEL_PATH,\n    trust_remote_code=True,\n    torch_dtype=\"auto\",\n    device_map=\"auto\",\n)\n\n# 2. Prepare Input\nmessages = [{\"role\": \"user\", \"content\": \"Explain the significance of the number 42.\"}]\ninputs = tokenizer.apply_chat_template(\n    messages,\n    tokenize=True,\n    add_generation_prompt=True,\n    return_dict=True,\n    return_tensors=\"pt\",\n).to(model.device)\n\n# 3. Generate\ngenerated_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)\noutput_text = tokenizer.decode(generated_ids[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)\n\nprint(output_text)\n```\n\n### 6.4 llama.cpp\n\n#### System Requirements\n- GGUF Model Weights(int4): 111.5 GB\n- Runtime Overhead: ~7 GB\n- Minimum VRAM: 120 GB (e.g., Mac studio, DGX-Spark, AMD Ryzen AI Max+ 395)\n- Recommended: 128GB unified memory\n#### Notes\n- llama.cpp build **must include** [PR #19283](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F19283)\n- If you encounter tool calling issues on llama.cpp mainline, try building with: [PR #18675](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F18675)\n- Community GGUF quantizations: [Step-3.5-Flash Quant Collection](https:\u002F\u002Fhuggingface.co\u002Fubergarm\u002FStep-3.5-Flash-GGUF#quant-collection)\n#### Steps\n1. Use official llama.cpp\n> the folder `Step-3.5-Flash\u002Ftree\u002Fmain\u002Fllama.cpp` is **obsolete**\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\ncd llama.cpp\n```\n2. Build llama.cpp on Mac:\n```bash\ncmake -S . -B build-macos \\\n  -DCMAKE_BUILD_TYPE=Release \\\n  -DGGML_METAL=ON \\\n  -DGGML_ACCELERATE=ON \\\n  -DLLAMA_BUILD_EXAMPLES=ON \\\n  -DLLAMA_BUILD_COMMON=ON \\\n  -DGGML_LTO=ON\ncmake --build build-macos -j8\n```\n3. Build llama.cpp on DGX-Spark:\n```bash\ncmake -S . -B build-cuda \\\n  -DCMAKE_BUILD_TYPE=Release \\\n  -DGGML_CUDA=ON \\\n  -DGGML_CUDA_GRAPHS=ON \\\n  -DLLAMA_CURL=OFF \\\n  -DLLAMA_BUILD_EXAMPLES=ON \\\n  -DLLAMA_BUILD_COMMON=ON\ncmake --build build-cuda -j8\n```\n4. Build llama.cpp on AMD Windows:\n```bash\ncmake -S . -B build-vulkan \\\n  -DCMAKE_BUILD_TYPE=Release \\\n  -DLLAMA_CURL=OFF \\\n  -DGGML_OPENMP=ON \\\n  -DGGML_VULKAN=ON\ncmake --build build-vulkan -j8\n```\n5. Run with llama-cli\n```bash\n.\u002Fllama-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?\"\n```\n6. Test performance with llama-batched-bench ([Benchmark details](llama.cpp\u002Fdocs\u002Fstep3.5-flash.md))\n```bash\n.\u002Fllama-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 1\n```\n\n## 7. Using Step 3.5 Flash on Agent Platforms\n\n> **Note:** As mentioned in the [Quick Start](#5-quick-start), you can access Step 3.5 Flash via **OpenRouter** or the **StepFun Platform**. Choose the base URL and API key corresponding to your preferred provider when configuring these agents.\n\n### 7.1 OpenClaw (Recommended)\n\n[OpenClaw](https:\u002F\u002Fopenclaw.ai) is a powerful agentic platform that works seamlessly with Step 3.5 Flash. \nStep 3.5 Flash is a perfect fit for OpenClaw due to its high speed and strong agentic capabilities (maintaining deep reasoning and consistency during execution).\n\n**Quick Setup:**\n1. **Install:** `curl -fsSL https:\u002F\u002Fopenclaw.ai\u002Finstall.sh | bash`\n2. **Onboard:** Run `openclaw onboard`.\n3. **Configure:** In WebUI (`Config` -> `Models`), add a new provider:\n   - **Type:** `openai-completions`\n   - **Base URL:**\n     - **StepFun:** `https:\u002F\u002Fapi.stepfun.ai\u002Fv1` (International) or `https:\u002F\u002Fapi.stepfun.com\u002Fv1` (China)\n     - **OpenRouter:** `https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1`\n   - **Model ID:** `step-3.5-flash` (or `stepfun\u002Fstep-3.5-flash` for OpenRouter)\n\nFor a full walkthrough, see our [OpenClaw Cookbook](.\u002Fcookbooks\u002Fopenclaw).\n\n### 7.2 Claude Code\n\nYou can configure [Claude Code](https:\u002F\u002Fdocs.anthropic.com\u002Fen\u002Fdocs\u002Fagents-and-tools\u002Fclaude-code\u002Foverview) to use Step 3.5 Flash by modifying your settings.\n\n**Update `~\u002F.claude\u002Fsettings.json`:**\n\n```json\n{\n  \"env\": {\n    \"ANTHROPIC_AUTH_TOKEN\": \"YOUR_STEPFUN_API_KEY\",\n    \"ANTHROPIC_BASE_URL\": \"https:\u002F\u002Fapi.stepfun.ai\u002F\"  \u002F\u002F Use \"https:\u002F\u002Fapi.stepfun.com\u002F\" for China endpoint\n  },\n  \"model\": \"step-3.5-flash\"\n}\n```\n\n> **Note:** For more advanced routing or OpenAI-style usage, we recommend [claude-code-router](https:\u002F\u002Fgithub.com\u002Fmusistudio\u002Fclaude-code-router).\n\n### 7.3 OpenAI Codex (CLI)\n\nTo use Step 3.5 Flash with the `@openai\u002Fcodex` CLI, update your `~\u002F.codex\u002Fconfig.toml`:\n\n```toml\nmodel=\"step-3.5-flash\"\nmodel_provider = \"stepfun-chat\"\npreferred_auth_method = \"apikey\"\n\n[model_providers.stepfun-chat]\nname = \"OpenAI using response\"\nbase_url = \"https:\u002F\u002Fapi.stepfun.ai\u002Fv1\"\nenv_key = \"OPENAI_API_KEY\"\nwire_api = \"chat\"\n```\n\n### 7.4 Step-DeepResearch\n\nStep 3.5 Flash powers the reasoning core of Step-DeepResearch.\nTo use it, simply set `MODEL_NAME` to `Step-3.5-Flash` in your `.env` file.\n\nSee the [Step-DeepResearch Repository](https:\u002F\u002Fgithub.com\u002Fstepfun-ai\u002FStepDeepResearch) for full setup instructions.\n\n\n## 8. Cookbooks\n\nExplore our [Cookbooks](.\u002Fcookbooks) directory for practical examples and integration guides, including:\n\n- **[Hybrid Local Agent](.\u002Fcookbooks\u002Fhybrid-local-agent-macos)**: Build a local, privacy-first agentic sandbox.\n- **[OpenClaw Integration](.\u002Fcookbooks\u002Fopenclaw)**: Deploy and integrate with OpenClaw.\n- **[Roo Code Integration](.\u002Fcookbooks\u002Froo-code-integration-guide)**: Use Step 3.5 Flash as the backend for Roo Code in VS Code.\n\nVisit the [Cookbooks README](.\u002Fcookbooks\u002FREADME.md) for more details and contribution guidelines.\n\n## 9. Known Issues and Future Directions\n\n1. **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.\n2. **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.\n3. **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.\n4. **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.\n\n## 10. Co-Developing the Future\n\nWe view our roadmap as a living document, evolving continuously based on real-world usage and developer feedback.\nAs 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.\n\n- **Join the Conversation**: Our Discord community is the primary hub for brainstorming future architectures, proposing capabilities, and getting early access updates 🚀\n- **Report Friction**: Encountering limitations? You can open an issue on GitHub or flag it directly in our Discord support channels.\n\n## License\nThis project is open-sourced under the [Apache 2.0 License](https:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0).\n\n## Star History\n\n\u003Ca href=\"https:\u002F\u002Fwww.star-history.com\u002F?repos=stepfun-ai%2FStep-3.5-Flash&type=date&legend=top-left\">\n \u003Cpicture>\n   \u003Csource media=\"(prefers-color-scheme: dark)\" srcset=\"https:\u002F\u002Fapi.star-history.com\u002Fimage?repos=stepfun-ai\u002FStep-3.5-Flash&type=date&theme=dark&legend=top-left\" \u002F>\n   \u003Csource media=\"(prefers-color-scheme: light)\" srcset=\"https:\u002F\u002Fapi.star-history.com\u002Fimage?repos=stepfun-ai\u002FStep-3.5-Flash&type=date&legend=top-left\" \u002F>\n   \u003Cimg alt=\"Star History Chart\" src=\"https:\u002F\u002Fapi.star-history.com\u002Fimage?repos=stepfun-ai\u002FStep-3.5-Flash&type=date&legend=top-left\" \u002F>\n \u003C\u002Fpicture>\n\u003C\u002Fa>\n","Step 3.5 Flash 是一个高性能、高效且可靠的开源基础模型，专为前沿推理和代理智能设计。该项目基于稀疏的混合专家（MoE）架构，能够在处理每个令牌时仅激活其196亿参数中的11亿，从而在保持实时交互所需敏捷性的同时，实现与顶级私有模型相媲美的推理深度。它支持多种应用场景，包括但不限于需要复杂决策支持和快速响应的环境。此外，Step 3.5 Flash 提供了详尽的文档和支持资源，便于用户快速上手并在本地或云端部署该模型。",2,"2026-06-11 03:52:25","high_star"]