StepFun Unveils Step 5 Preview: 600B MoE, 27B Active, Open-Source Top 2
StepFun released Step 5 Preview, a 600B-parameter MoE model with 27B active parameters and 1M context, available as an open-weight checkpoint on Hugging Face. In Artificial Analysis evaluations, it scored 44, ranking as the top open-source model after one other, with pricing at $1 per million input tokens and $2.7 per million output tokens. Hands-on tests showed strong agentic capabilities, including autonomously building 3D assets and a LEGO racing game.
Coverage timeline
StepFun 模型发布 (HF)
stepfun-ai/Step-5-Preview-BF16 Updated 10 minutes ago • 2
量子位量子位
开源Top2!实测阶跃Step 5 Preview,真有点猛啊… Jay 2026-09-21 11:46:43 来源: 量子位 激活参数仅27B Jay 发自 凹非寺 量子位 | 公众号 QbitAI 谁能想到,全球旗舰模型混战打到现在,局势依旧充满变数。 昨天, 阶跃星辰 毫无预兆地端出了最新一代旗舰基座模型——Step 5 Preview。 MoE架构,600B总参数、激活参数仅27B,1M上下文。 Agent能力大幅提升,在Artificial Analysis的最新评测中,Step 5 Preview取得44分,直接冲到全球开源Top 2。 要知道,其他拿到这一分数的大模型,大多都是万亿参数级别。 所以相比起来,阶跃这个新模型定价蛮便宜。 百万输入:1美元 百万输出:2.7美元 可以说很有竞争力了,单个 任务成本仅为Opus 5的12.5% 。 下面这张图更加直观。 AA榜单中, Step 5 Preview位于Intelligence Index与单任务成本权衡的「帕累托前沿」 。 说实话,看到这个消息时我有点意外。 阶跃挺久没出现在这么靠前的位置了,之前两代模型都聚焦在Flash这个档位。 但Benchmark这东西嘛,现在大家也都懂。 榜单上你追我赶,今天刷掉一个,明天又冒出来一个。真到要用的时候,模型到底行不行,完全是另外一码事。 所以我干了一件比较费钱的事: 烧了无数Token,并行跑了无数个3D资产、3D游戏…… 然后,我发现—— 这模型好像还真可以啊。。。 一手实测 最近Astra操控软件不是很火吗? 所以API一接上,我直接给Step 5 Preview上强度,让它操作Blender,给我建一个后室出来。 说实话,最开始我没指望它真能交付出什么东西。 结果等它自己折腾了一个多小时,我打开文件夹,再点开渲染好的MP4,直接看愣了。 不er,这啥啊,你搁哪个网站下载的视频?? 建模本身就不说了, 离谱的是,它居然自己给视频加了一堆后期 。 VHS录像质感、镜头噪点、昏黄灯光,人物走路的时候甚至还有脚步声。 给我看得有点毛骨悚然了,感觉转角会跳出一个《痴迷》女主。。。 所以咱还是做点老少皆宜的demo吧。 第二个任务,我让它照着1999年的《LEGO Racers》,直接搓一个乐高赛车游戏。 这次也挺完整。 赛道、赛车、人机车手、实时排名都有,甚

Related stories
Hugging Face Details 2.8T-Token Indic Pretraining with Open Data Pipeline
Hugging Face has published details of a large-scale Indic pretraining effort involving 2.8 trillion tokens, alongside an open data generation pipeline. The initiative aims to improve language model performance for Indian languages by providing transparent, reproducible data creation methods.
1 reports · 1 sources
Meta releases Muse Spark Open, an 816B-parameter open-weight model optimized for Apple silicon
Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model with 816B total and 42B active parameters and a 1M-token context window, claiming it is the most capable US open-weight model. Optimized for Apple silicon via MLX, it can run locally on two M5 Ultra Mac Studios, with day-0 support from vLLM and other inference engines.
6 reports · 6 sources

StepFun Releases 321B-Parameter Quantized Step-5-Preview-NVFP4 on Hugging Face
StepFun has released Step-5-Preview-NVFP4, a 321B-parameter quantized variant of its Step-5-Preview model, on Hugging Face. The model is available under the stepfun-ai organization and was updated 36 minutes ago.
1 reports · 1 sources
StepFun Releases Step-5-Preview, a 604B-Parameter Open-Weight LLM
StepFun has released Step-5-Preview, a 604B-parameter open-weight large language model, on Hugging Face. The model is available under the stepfun-ai/Step-5-Preview repository, updated 35 minutes ago.
1 reports · 1 sources
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe