Step 3.5 Flash - The latest open-source base model from Step Star.

Step 3.5 What is Flash?

Step 3.5 Flash is StepStar's latest open-source foundation model, specifically designed for Agent scenarios. The model employs a sparse MoE architecture with a total of 196 billion parameters, activating only 11 billion parameters per token, balancing performance and efficiency. Step 3.5 Flash boasts an inference speed of up to 350 TPS, supports 256K long contexts, and rivals top-tier closed-source models in mathematical inference, code generation (SWE-bench 74.4%), and Agent tasks. Step 3.5 Flash is open-source and supports frameworks such as vLLM, SGLang, and llama.cpp, enabling local deployment on consumer-grade hardware like Mac Studio M4 Max and NVIDIA DGX Spark, achieving a balance between data privacy and high performance.

Step 3.5 Main Functions of Flash

• High-speed reasoningThe model achieves a generation speed of up to 350 TPS through MTP-3 technology, supporting real-time response for complex multi-step inference.
• Agent capabilitiesThe model is designed specifically for agent tasks, achieving a 74.4% accuracy rate on the SWE-bench Verified, and can handle complex tasks with long chains of operations.
• High-efficiency long textSupports 256K context windows and employs a hybrid attention mechanism to reduce computational overhead for long texts.
• Local deploymentOptimized for consumer hardware support, it can run smoothly on devices such as Mac Studio M4 Max and NVIDIA DGX Spark.
• Code generationThe model has powerful programming capabilities, supporting automatic tool invocation and structured reasoning output.

Step 3.5 Technical Principles of Flash

• Sparse MoE architectureThe model employs a 45-layer Transformer backbone network, with each layer configured with 288 fine-grained routing experts and 1 shared expert. During inference, only the Top-8 experts are activated, and approximately 11 billion parameters are actually computed per token, achieving a balance between the model's capacity of 196 billion total parameters and the inference cost of the small model.
• MTP-3 Multi-Token PredictionBy employing a dedicated prediction head composed of a sliding window attention mechanism and a dense feedforward network, four tokens are generated in parallel during a single forward propagation. This increases the generation speed for typical scenarios to 100-300 tok/s, with a peak speed of 350 tok/s, significantly reducing decoding latency.
• Hybrid attention mechanismThe architecture employs a 3:1 alternating sliding window attention layer and global attention layer. The sliding window layer focuses on local context, while the global layer captures long-distance dependencies, effectively controlling computational complexity and balancing efficiency and performance in scenarios with 256K long texts.
• Inference optimization strategyThe model supports a combined deployment of expert parallelism (EP8) and tensor parallelism (TP8), with FP8 quantization reducing memory bandwidth pressure. Through speculative decoding and MTP collaboration, it achieves efficient service-oriented deployment on Hopper GPUs.

Step 3.5 Flash project address

• GitHub repositoryhttps://github.com/stepfun-ai/Step-3.5-Flash/
• HuggingFace model libraryhttps://huggingface.co/stepfun-ai/Step-3.5-Flash

Step 3.5 Application Scenarios of Flash

• Intelligent programming developmentAs the underlying model for tools such as Claude Code and Codex, it provides capabilities such as code generation, automatic debugging, and software engineering task processing, achieving a pass rate of 74.4% on SWE-bench Verified.
• Autonomous agent executionIt is suitable for agent scenarios that require long-chain reasoning, such as in-depth research, web information retrieval, and cross-platform data comparison.
• Real-time dialogue interactionWith a generation speed of 100-350 TPS, it supports interactive applications that require real-time response, such as low-latency chatbots, online education and tutoring, and intelligent customer service.
• Long text analysis and processingIt can be used for academic paper reading, legal contract review, and understanding of large code libraries, efficiently extracting and integrating massive amounts of information.
• End-side privacy computingIt can be deployed on local devices such as Mac Studio M4 Max and NVIDIA DGX Spark to meet the private processing needs of sensitive data in finance, healthcare, and enterprise office.