StepFun releases the open Step 3 multimodal reasoning model

Background The World Artificial Intelligence Conference (WAIC, 世界人工智能大会) is a major AI conference. StepFun (阶跃星辰) chose the event to release Step 3 ^[3].

What happened StepFun released Step 3 in July 2025, during WAIC ^[3]. Its system technical report, "Step-3 is Large yet Affordable", was posted to arXiv on 25 July 2025 ^[1].

Step 3 has 321 billion total parameters, with 38 billion activated per token ^[1]. The weights were released under Apache 2.0. Chinese Wikipedia dates the open-source release to 31 July 2025 ^[2][4].

Key details StepFun reported decoding throughput of up to 4,039 tokens per second per GPU. That is about 1.7 times the 2,324 tokens per second that DeepSeek-V3 achieved in the same setup on Hopper GPUs ^[1].

At WAIC, StepFun formed a Model-Chip Ecosystem Innovation Alliance with Huawei (华为), Biren (壁仞科技), Moore Threads (摩尔线程) and Enflame (燧原科技). The alliance aims to optimise Step 3 for domestic chips ^[3].

Significance The release combined an open reasoning model with a plan to run it on Chinese-made chips. The throughput figures are StepFun's own, reported in its technical report.

Terms used A reasoning model is designed to work through a problem in steps before giving an answer. Decoding throughput is the number of output tokens a system can generate per second, and it largely determines the cost of serving a model: more tokens per second on the same hardware means more answers served. A time-per-output-token limit sets a maximum delay between successive tokens, which matters for interactive use.

Why chip optimisation matters A model's speed depends on the hardware it runs on. Tuning a model to a particular family of chips can change how fast it runs and how much it costs to serve on that chip, which is the purpose of the alliance with the chip makers.