阶跃星辰（StepFun）发布 Step 5 Preview 开放权重，Sparse MoE 架构总参数 600B、每 token 激活 27B，支持 1M token 上下文窗口和原生文本与视觉输入，主打长时程智能体任务如编码、软件工程和金融。
NVIDIA 补充了 GB300 上的推理吞吐和部署路径，读者可据此评估该开源权重模型的实际落地方式。
Congrats to StepFun on Step 5 Preview 🙌 Their largest open weights release.
The 600B MoE model is built to handle agent tasks lasting up to 24 hours.
The model runs at nearly 8K tokens/sec/GPU on NVIDIA GB300 using NVFP4.
NVP4 checkpoint on @huggingface: https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4
Deploy with @SGL_project - https://docs.sglang.io/cookbook/autoregressive/StepFun/Step-5-Preview or @vLLM - https://recipes.vllm.ai/stepfun-ai/Step-5-Preview
Post-train with NeMo Automodel: https://github.com/NVIDIA-NeMo/Automodel/blob/main/docs/model-coverage/llm/stepfun-ai/step-5-preview.mdx
Run it with NVIDIA NIM using a single command: https://catalog.ngc.nvidia.com/orgs/nim/nvidia/containers/vllm-model-free-nim-experimental/-
Step 5 Preview is now open weights. BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog. • Sparse MoE: 600B total, 27B active per token • 1M-token context window • Native text and vision input • Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance Frontier intelligence, open weights. Build what's next. Hugging Face: https://huggingface.co/stepfun-ai/Step-5-Preview ModelScope: https://modelscope.ai/models/stepfun-ai/Step-5-Preview GitHub: https://github.com/stepfun-ai/Step-5-Preview在 X 查看被引用的帖子
来源：NVIDIA AI · x.com