StepFun

StepFun's new flagship for agentic work, announced September 20, 2026 under the banner "Advancing the Pareto Frontier": the pitch is frontier-adjacent…

StepFun's audio-language foundation model for spoken interaction, organised as a continuous listen–converse–think–act loop. Deep Per…

StepFun's long-form music generation model (with ACE Studio), released through the StepFun API with a technical report on 11 September 2026. The desig…

StepFun's new flagship open-weight reasoning VLM. 198B total / ~11B active sparse MoE with 256K context, tunable reasoning tiers, and ~400 tok/s outpu…

5-dimensional evaluation benchmark for GUI and web agents.

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…

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

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

Research on parallel coordinated reasoning to improve agentic accuracy.

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

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

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…

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

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

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

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

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

Specialized evaluation set for competitive programming and math reasoning.

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

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

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

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

StepFun and Tsinghua (first author Jingcheng Hu; senior author Harry Shum) delivered the first fully open implementation of large-scale reasoning-orie…

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

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

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

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

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

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