Step-3.5-Flash

What is Step-3.5-Flash?

Step-3.5-Flash is a mixture-of-experts model with 196B parameters, of which only 11B work on each token, so it serves quickly and cheaply. It reads and writes text, holds 256K tokens of context and calls tools through an OpenAI-compatible API. The weights are free under Apache 2.0, and the 4-bit build needs about 120 GB of memory.

StepFun still serves it, next to an agent-tuned step-3.5-flash-2603 version at the same price. Its newer Step 3.7 Flash builds on it with image and video input at a higher token price.

Step-3.5-Flash or Step 3.7 Flash?

Step 3.7 Flash builds on Step-3.5-Flash and adds image and video input; Step-3.5-Flash stays the cheaper pick for text-only coding and agent work.

Input

Step-3.5-Flash reads text only; Step 3.7 Flash also understands images and video.

Price

$0.10 / $0.30 per million tokens against $0.20 / $1.15 for Step 3.7 Flash, nearly four times as much for output.

Size and context

Both activate 11B parameters per token (196B against 198B in total), and both hold 256K tokens.

Weights

Both are open under Apache 2.0 on Hugging Face, and both are in every Step Plan tier.

Who is Step-3.5-Flash best for?

Step-3.5-Flash is best for developers who want a low-cost open-weight model for coding agents and tool calling, through StepFun's API or on a 128 GB machine of their own.

• Coding agents

  StepFun reports 74.4% on SWE-bench Verified, and its Step Plan connects the model to Claude Code, Cursor, OpenClaw and similar tools.

• High-volume agent calls

  Only 11B of its 196B parameters are active per token, and StepFun quotes 100 to 300 tokens a second in typical use.

• Long codebases and documents

  A 256K-token context holds large repositories or long reports in a single request.

• Running a model yourself

  The 4-bit weights need about 120 GB of memory, so a 128 GB Mac Studio, NVIDIA DGX Spark or AMD Ryzen AI Max+ 395 machine can run it.

What's included with Step-3.5-Flash?

Apache 2.0 weights, serving through vLLM, SGLang and llama.cpp, fast multi-token decoding, an agent-tuned version and a Step Plan for coding tools.

Open weights

Apache 2.0 weights on Hugging Face and ModelScope, with FP8 and GGUF builds plus the base and mid-training checkpoints.

Runs on common servers

StepFun documents vLLM, SGLang, Hugging Face Transformers and llama.cpp, with a tool-call parser for agent use.

Multi-token prediction

It predicts three tokens at a time, which StepFun says gives 100 to 300 tokens a second and up to 350 on single-stream coding.

Agent-tuned 2603 version

Same price, fewer tokens per task, and a low reasoning setting through the reasoning_effort field.

Step Plan for coding tools

Monthly credits that work in Claude Code, Cursor, OpenClaw and Trae, from $6.99 a month.

How much does Step-3.5-Flash cost?

Step-3.5-Flash costs $0.10 per million input tokens and $0.30 per million output tokens on StepFun's API, Step Plan subscriptions start at $6.99 a month, and the open weights are free to download.

Flash Max costs $99 a month for 40,000M credits. Paid quarterly, Mini, Plus, Pro and Max cost $18.99, $26.99, $79 and $269; paid yearly, $69.99, $95.99, $289 and $989. StepFun values about 7M credits at $1 of model use; credits arrive monthly, expire at month end, and booster packs cost $6.99 for 400M or $9.99 for 1,600M. The agent-tuned step-3.5-flash-2603 costs the same per token. The weights are free under the Apache 2.0 licence on Hugging Face and ModelScope. Full pricing on platform.stepfun.ai. We checked these prices on 28 Sep 2026.

How do readers rate Step-3.5-Flash?

No one has rated Step-3.5-Flash yet. Be the first.

Step-3.5-Flash pros and cons

It is cheap, fast and strong for its size, but it spends many tokens thinking, knows less than larger models and reads text only.

Is Step-3.5-Flash open source?

Yes. Step-3.5-Flash's code is public under the Apache 2.0 licence, so you can run it on your own server.

What are the best Step-3.5-Flash alternatives?

Step-3.5-Flash is an open source alternative to MiniMax M2.5 and Gemini 3.8 Flash. DeepSeek-V3.2, DeepSeek V4.1 Flash and MiMo-V2.6-Flash do the same job and are open source too.

MiniMax M2.5Free

MiniMax's open-weight coding and agent model, at $0.30 per million input tokens.

Gemini 3.8 FlashFree

Google's Gemini 3.8 Flash for coding, reasoning and agents.

DeepSeek-V3.2Open Source

DeepSeek's open, MIT-licensed reasoning model for agents, with thinking during tool calls.

DeepSeek V4.1 FlashOpen Source

Choose it when you need image input or a 1M-token context; its API costs $0.30 / $1.20 per million tokens at peak hours and half that off-peak.

MiMo-V2.6-FlashOpen Source

Xiaomi's 309B open omnimodal model for coding and agents, MIT license.

Laguna S 2.1Open Source

Choose it when you want a 1M-token context on a smaller model: 118B parameters with 8B active, a free OpenRouter endpoint, and a 1M-context endpoint at $0.10 / $0.20.

Questions about Step-3.5-Flash