This is a simplified guide to an AI model called Step-3.5-Flash maintained by stepfun-ai. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter. This is a simplified guide to an AI model called Step-3.5-Flash Step-3.5-Flash maintained by stepfun-ai stepfun-ai . If you like these kinds of analysis, join AIModels.fyi AIModels.fyi or follow us on Twitter Twitter . Model overview Step-3.5-Flash is an open-source foundation model from stepfun-ai engineered for frontier reasoning and agentic capabilities. Built on a sparse Mixture of Experts architecture, it activates only 11B of its 196B parameters per token, making it far more efficient than comparable models. This intelligent parameter allocation allows it to match the reasoning depth of top-tier proprietary systems while maintaining rapid inference speeds. Compared to similar models like step3 and deepseek-v3, this variant focuses on speed and accessibility rather than multimodal capabilities. Step-3.5-Flash stepfun-ai step3 deepseek-v3 Model inputs and outputs Step-3.5-Flash accepts text prompts and can engage in complex reasoning tasks, code generation, and agentic workflows. The model processes queries through a 45-layer transformer with a 256K context window, allowing it to handle lengthy documents and codebases. It outputs coherent text responses with support for multi-step reasoning and technical problem-solving. Step-3.5-Flash Inputs Text prompts of any length up to 256K tokensConversation history for maintaining context across multiple exchangesCode snippets or technical specifications for analysis and generationComplex multi-part questions requiring reasoning chains Text prompts of any length up to 256K tokens Text prompts Conversation history for maintaining context across multiple exchanges Conversation history Code snippets or technical specifications for analysis and generation Code snippets or technical specifications Complex multi-part questions requiring reasoning chains Complex multi-part questions Outputs Reasoned text responses with step-by-step explanationsGenerated code for software engineering tasksAgent-driven actions including web browsing and task executionMathematical proofs and solutions with detailed working Reasoned text responses with step-by-step explanations Reasoned text responses Generated code for software engineering tasks Generated code Agent-driven actions including web browsing and task execution Agent-driven actions Mathematical proofs and solutions with detailed working Mathematical proofs and solutions Capabilities The model generates text at 100-300 tokens per second in typical usage, with peaks reaching 350 tokens per second for single-stream coding tasks. This speed comes from its 3-way Multi-Token Prediction head, which predicts four tokens simultaneously in a single forward pass. For code-related work, it achieves 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0, demonstrating strong performance on long-horizon software engineering tasks. The hybrid attention mechanism—combining three Sliding Window Attention layers with full-attention layers—maintains consistent performance across the entire 256K context window without proportional increases in computational cost. What can I use it for? Step-3.5-Flash serves software developers building agentic systems that require rapid decision-making and code synthesis. Companies can deploy it locally for privacy-critical applications on high-end consumer hardware like Mac Studio M4 Max or NVIDIA systems. Research teams benefit from the open architecture for extending or fine-tuning reasoning capabilities. The model handles complex information retrieval tasks through its agentic abilities, scoring 88.2 on Agency-Bench and 83.7 on xbench-DeepSearch. You can integrate it via cloud APIs through OpenRouter or the StepFun platform for straightforward API access without managing infrastructure yourself. Step-3.5-Flash OpenRouter StepFun platform Things to try Test the model's reasoning on competition-level mathematics—it scores 97.3% on AIME 2025 and 98.4% on HMMT 2025, making it suitable for tutoring systems or automated grading. Try deploying it as a local coding copilot for refactoring large legacy codebases, leveraging its efficient context handling to process entire files without token budget constraints. Experiment with agentic workflows where the model autonomously browses the web, executes terminal commands, and manages complex research tasks. The model's sparse activation pattern means you can run sophisticated reasoning on modest hardware, enabling personal AI projects that would otherwise require expensive cloud compute.

A 196B Open Model That Runs Like It’s 11B: Step-3.5-Flash’s Trick

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

10 Languages, 9 Premium Voices: Meet Qwen3-TTS CustomVoice

The 5 Principles Every AI Builder Needs to Master in 2025

Beyond Keywords: Engineering a Production-Ready Agentic Search Framework in Go

From Copilot to Coworker: Moving Beyond "Autocomplete" to "Autonomous Agents" in the IDE

Emotion Is Everywhere in AI—Except in the System That Runs It

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

10 Languages, 9 Premium Voices: Meet Qwen3-TTS CustomVoice

The 5 Principles Every AI Builder Needs to Master in 2025

Beyond Keywords: Engineering a Production-Ready Agentic Search Framework in Go

From Copilot to Coworker: Moving Beyond "Autocomplete" to "Autonomous Agents" in the IDE

Emotion Is Everywhere in AI—Except in the System That Runs It

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps