vLLM and SGLang release self-hosting recipes for StepFun Step 5 Preview
Two Inference Stacks Release Local Deployment Recipes for Step 5 Preview
On October 15, 2026, self-hosting infrastructure for large language models shifted as vLLM and SGLang independently published deployment stacks for StepFun’s Step 5 Preview model, as reported by OrcaRouter. Until this release, engineering teams wanting to run the flagship model on proprietary hardware relied entirely on community re-uploads of weights carrying uncertain licensing terms.
Software Artifacts and Open Pull Requests Released on October 15
The October 15 infrastructure drop comprised four distinct technical artifacts carrying identical timestamps. In the vLLM ecosystem, contributor jeejeelee opened Pull Request #61727 titled “Add Step-5 Preview support” at 01:18 UTC as a draft commit spanning 129 files and 16,996 lines of code. Simultaneously, a corresponding configuration recipe appeared under `vllm-project/recipes` in Pull Request #1122, introducing `models/stepfun-ai/Step-5-Preview.yaml` with explicit FP8 and NVFP4 parameter variants. A companion container image designated as `vllm/vllm-openai:step-5-preview` landed on Docker Hub at 00:51 UTC, providing dual-architecture manifests for `linux/amd64` (8.56 GB compressed) and `linux/arm64` (9.58 GB compressed), alongside explicit CUDA 13.0 targets (`cu130`). Meanwhile, SGLang merged Pull Request #44357 into its main branch at 01:21 UTC, establishing a dedicated cookbook page at `docs.sglang.ai/cookbook/autoregressive/StepFun/Step-5-Preview`. The accompanying SGLang support pull request, #44356, remained open, with the documentation explicitly noting that standard `pip install sglang` commands cannot yet instantiate the model because support has not shipped in a stable release.
Sparse Mixture-of-Experts Architecture and Parameter Footprint
Analysis of the newly published serving loaders reveals the internal mechanics of Step 5 Preview, a model featuring a one-million-token context window, multi-modal input processing for text, images, and video, and text-only generation capped at a 64k-token output limit. According to the first-party serving documents analyzed by OrcaRouter, the model utilizes a sparse Mixture-of-Experts architecture containing 604 billion total parameters, of which roughly 27 billion remain active per token. The network spans 92 layers, where each MoE layer routes individual tokens across 8 out of 352 available experts alongside a single shared expert. This expert count required vLLM to implement a custom router GEMM kernel modification supporting a `(4096, 352)` hidden-dimension and expert pair. For speculative decoding, the architecture integrates three Multi-Token Prediction (MTP) layers.
Hardware Memory Baselines and Blackwell Versus Hopper Constraints
Deploying the model locally requires substantial hardware acceleration, with vLLM and SGLang recipes defining strict VRAM floors. The FP8 parameter configuration demands a minimum of 773 GB of VRAM distributed across at least four GPUs, targeting NVIDIA Blackwell B300 hardware configurations utilizing Tensor-Parallel (TP4), Expert-Parallel (TEP4), or Data-Parallel (DEP4) layouts. SGLang documentation notes that initial compilation runs on B300 hardware consume approximately 15 minutes via FlashInfer, dropping to six minutes on subsequent boots provided that workspace caches remain persistent.
API Economics and Independent Benchmark Performance
Commercial API access maintained by StepFun lists input pricing at $1.00 per million tokens for cache misses, dropping to $0.10 per million tokens for cache hits, with output generation priced at $2.70 per million tokens. Independent evaluation by Artificial Analysis assigns the model an Intelligence Index score of 43.7, paired with a median output velocity of 88.5 tokens per second and a time-to-first-token latency of 24.7 seconds. These latency metrics reflect the model’s internal reasoning behavior, which invariably executes a thinking phase prior to output generation—generating roughly 2,000 reasoning tokens per query.
When Was Step 5 Preview Released and How Do vLLM and SGLang Differ?
When was Step 5 Preview originally made available to developers? StepFun made the model accessible via its proprietary API on September 20, 2026, though independent evaluator Artificial Analysis began tracking performance data on September 18, 2026.
What is the primary architectural difference between vLLM and SGLang handling of this model? The vLLM recipe explicitly notes that Simple and LMCache key-value offloading are not supported for Step 5 Preview, whereas SGLang documents its HiCache path as fully available in the newly published cookbook.
Disclaimer: The technical analyses and security protocols detailed in this article are for informational purposes only. Always consult with certified IT and cybersecurity professionals before altering enterprise networks or handling sensitive data.