Step 3.5 Flash

Step 3.5 Flash is StepFun's 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, achieving high intelligence density. This makes it ideal for developers, researchers, and agent builders who require deep reasoning without sacrificing speed. The model's design prioritizes inference cost and speed as core constraints, enabling real-time interaction while rivaling top-tier proprietary models. Its core value lies in combining extensive parametric knowledge with agile per-token computation, making it suitable for a wide range of advanced AI applications.

The primary pain point addressed by Step 3.5 Flash is the trade-off between model capability and inference efficiency in large language models. Traditional dense models require enormous computational resources for every query, making them impractical for real-time or interactive use. Agentic tasks demand both deep reasoning and immediate responsiveness, which many models struggle to balance. Step 3.5 Flash solves this by decoupling total knowledge capacity from per-token computation, activating only a fraction of its parameters while maintaining frontier-level performance. This allows users to deploy a powerful AI system without excessive hardware costs or latency, enabling agents to reason and act fluidly in dynamic environments.

The first major feature group is the Sparse Mixture of Experts backbone and 3-way Multi-Token Prediction (MTP-3). The MoE architecture spans 196B total parameters but only activates 11B per token, drastically reducing memory and computational overhead. MTP-3 enables the model to predict multiple tokens simultaneously, achieving generation throughput of 100-300 tok/s in typical usage, peaking at 350 tok/s for single-stream coding tasks. This combination allows for complex, multi-step reasoning chains with immediate responsiveness, a critical capability for agentic workflows where speed and depth are equally important. The result is an agent that can think fast without sacrificing accuracy.

The second major feature group focuses on coding and agentic performance through a scalable Reinforcement Learning framework. Step 3.5 Flash achieves 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0, demonstrating robust ability in handling sophisticated, long-horizon tasks with unwavering stability. The model integrates Python code execution within its Chain-of-Thought reasoning, boosting tool-augmented reasoning performance on benchmarks like AIME 2025 (99.8) and IMOAnswerBench (86.7). This dual capability allows the model to act as a professional data scientist or software engineer, performing end-to-end data analysis and code generation with high accuracy and reliability.