StepFun

startup

china · Founded 2023-04 · Valuation: $3B as of 2026-01

StepFun is one of China's "six AI tigers," founded in April 2023 in Beijing. Distinguished by its multimodal-first approach from inception, the company draws its core vision expertise from Megvii (Face++) — Chairman Yin Qi is Megvii's co-founder and CEO, and Chief Scientist Zhang Xiangyu is famous for co-authoring ResNet and ShuffleNet at Megvii Research.

Founder and CEO Jiang Daxin previously served as Corporate Vice President at Microsoft, leading Asia-Pacific internet engineering. Advisor Heung-Yeung Shum is a former Executive Vice President of Microsoft and former head of Microsoft Research Asia (MSRA), often called the "Whampoa Military Academy" of Chinese AI for having trained a generation of the country's top AI leaders.

StepFun has raised over $100M across multiple rounds from investors including Tencent, Xiaomi, Lenovo Capital, and Source Code Capital, reaching an estimated ~$2B valuation by mid-2024. The Step model series has iterated rapidly from Step-1 through Step-3.5-Flash, spanning language, vision, video, and audio modalities.

Strategically, StepFun leads a Model-Chip Alliance with Huawei, Biren, Moore Threads, and Enflame to validate models on domestic Ascend chips, addressing China's compute sovereignty challenges under US export controls. The company is also the primary AI provider for Geely Auto's Agent OS in mass-produced vehicles, powering multimodal in-cabin AI assistants.

Website Wikipedia HuggingFace GitHub Artificial Analysis
frontieropen-weightmultimodalmoereasoningvideoaudio

People

News

Outputs (24)

GEBench

dataset

5-dimensional evaluation benchmark for GUI and web agents.

Step-3.5-Flash 3

model paper dataset

Most advanced agentic model. 196B MoE (11B active) using MTP-3 and hybrid attention. Optimized for speed (350 tok/s) with 256k context. Step-3.5-Flash…

↓ 85.1k

ACE-Step-1.5

model

Open-source music foundation model generating full-length, high-fidelity songs on consumer hardware (under 4GB VRAM).

Step-3-VL-10B 2

model paper

"Compact giant" outperforming models 20x its size (including GPT-4o on certain benchmarks) via fully unfrozen perception-decoder training.

↓ 180.9k

PaCoRe: Learning to Scale Test-Time Compute

paper

Research on parallel coordinated reasoning to improve agentic accuracy.

↓ 530

Step-DeepResearch

library

Deep research agent framework for autonomous multi-step research tasks.

gelab-zero (STEP-GUI)

library

High-performance GUI agent for autonomous computer/phone use.

↓ 436

Step-Audio-R1

model

First audio LLM to unlock test-time compute scaling via Chain-of-Thought reasoning. 33B parameters. Surpasses Gemini 2.5 Pro on audio understanding be…

↓ 33

Step-Audio-EditX

model

3B parameter RL-based model for editing emotion, style, and paralinguistics in audio.

↓ 16.9k

NextStep-1 2

model paper

14B autoregressive image generation model that rivals top diffusion systems. ICLR 2026 Oral. Includes NextStep-1-Edit for instruction-based image edit…

↓ 37

StepFun-Prover-Preview

model

Specialized 32B model family for formal theorem proving (Lean 4), using tool-integrated reinforcement learning.

↓ 47

Step-3 2

model paper

321B parameter MoE model (38B active) designed for high-efficiency multimodal reasoning with cost-effective decoding.

↓ 73.4k

Predictable Scale Part II: Farseer

paper

Refined scaling law for predicting LLM training loss across scales. Introduces a loss surface function L(N,D) that reduces extrapolation error by 433%…

CF-Div2-Stepfun

dataset

Specialized evaluation set for competitive programming and math reasoning.

SteptronOss

library

Lightweight, AI-native training framework for LLMs (SFT, RLVR).

Step1X-3D: High-Fidelity Textured 3D Assets

model

High-fidelity and controllable generation of textured 3D assets.

Step1X-Edit

model

SOTA open-source instruction-based image editing model using MLLM + DiT architecture. Comparable to GPT-4o and Gemini 2 Flash on editing tasks.

↓ 62

Step-R1-V-Mini

model

Multimodal reasoning model with high-precision image perception. Trained via PPO reinforcement learning with verifiable rewards in the image space. Ra…

Predictable Scale Part I: Step Law

paper

Empirical study across 3,700+ LLMs trained on 100T tokens establishing optimal hyperparameter scaling laws. Finds optimal learning rate follows a powe…

Step-Audio / Step-Audio2

model

Multimodal audio models supporting 16.7Hz semantic tokenization and real-time "omni" interaction (speech-to-speech).

Step-Video-T2V 2

model paper

Massive 300B parameter text-to-video model capable of generating high-fidelity 204-frame clips.

Step-2

model

First trillion-parameter MoE model from a Chinese startup. Ranks among the top globally for instruction-following and reasoning.

Step-1

model

Debut 130B parameter LLM that established StepFun as a top-tier contender.

Step-1V / 1.5V / 2V

model

Series of vision-language models capable of understanding images and complex documents.