VLLM launches Day-0 support for StepFun's Step 5 Preview model

vLLM released full Day-0 support yesterday for StepFun-AI's Step 5 Preview, a multimodal MoE model that ships with a one-million-token context window, a hybrid attention mechanism combining sparse and sliding-window patterns, and MTP. The model is available in two variants: a full BF16 build and a quantized NVFP4 build that uses mixed-precision layers — some NVFP4, some BF16 — with the quantization scheme auto-detected from the configuration. Both variants run on the same Docker image, vllm/vllm-openai:step-5-preview.

Running the model requires access on Hugging Face, including hf auth login, and either four B300 (Blackwell) GPUs for an FP8 configuration or eight H200 (Hopper) GPUs for the same purpose. On Blackwell, vLLM offers three parallel deployment modes: TP4, TEP4, and DEP4. The latter two use MegaMoE for routed experts, while TP4 stays on the default backend because MegaMoE requires expert parallelism. On Hopper, the GPU count must be doubled to eight and the MegaMoE flag omitted. The --max-model-len parameter for the NVFP4 variant is capped at 320,000 tokens due to KV cache memory constraints, and the --reasoning-parser step5 and --tool-call-parser step5 flags remain fixed across both variants.

The more significant addition for developers is immediate reinforcement-learning support via verl, ByteDance's distributed training library. The recipe enables training the text model in FP8 with GRPO and LoRA on top of FSDP2, using the identical Docker image for both inference and training. This marks the first time vLLM has paired official RL support with an external model launch rather than limiting it to internal community models.