Step 5 Preview Is Self-Hostable on Day Zero: vLLM and SGLang Both Land Serving Paths for StepFun's 600B MoE

Step 5 Preview was not new on the morning of October 15, 2026. StepFun's 600-billion-parameter flagship has been callable through the vendor's own API since September 20, 2026, and the independent evaluator Artificial Analysis has been scoring it since September 18. What was new that morning was narrower and considerably more useful: two mainstream inference stacks published their own paths for running it on hardware you own or rent. vLLM opened a draft pull request adding Step 5 Preview — 129 files, +16,996 lines — alongside a companion recipe and a dedicated vllm/vllm-openai:step-5-preview image that landed on Docker Hub at 00:51 UTC. SGLang merged a cookbook page with Blackwell and Hopper deployment recipes, with the model-support pull request behind it opened the same hour. Until that morning, the only way to get Step 5 Preview onto your own GPUs was a community re-upload of a checkpoint nobody had licensed you to redistribute.

So the honest headline is not "StepFun shipped a model." It shipped this one in September, and the model's own date belongs in the first sentence of any piece about it. The headline is that as of October 15, 2026, the serving question has commands in the answer.

What actually changed on October 15

Four artifacts, all dated the same day, and they are worth separating because they carry different levels of commitment.

• vLLM model support — pull request #61727, titled "Add Step-5 Preview support", opened by jeejeelee at 01:18 UTC and carrying the draft flag. It is a single squashed commit touching 129 files, and it is not merged. A draft PR is an author asking for review, not a maintainer promise.

• The vLLM recipe — pull request #1122 in vllm-project/recipes, opened two minutes earlier at 01:16 UTC and also open, adds models/stepfun-ai/Step-5-Preview.yaml with a FP8 default variant and an NVFP4 variant. This is the document with the numbers in it, and it is the one most of this article draws on.

• The container image — vllm/vllm-openai:step-5-preview, pushed 2026-10-15 at 00:51 UTC as a two-architecture manifest: linux/amd64 at 8.56 GB compressed and linux/arm64 at 9.58 GB, with step-5-preview-x86_64-cu130 and step-5-preview-arm64-cu130 alongside them. A published image is a stronger signal than an open PR: it means somebody built this and pushed it, regardless of when the review finishes.

• The SGLang cookbook — pull request #44357, "Add Step 5 Preview cookbook", merged into main at 01:21 UTC after review, adding a deployment page the SGLang docs now serve at docs.sglang.ai/cookbook/autoregressive/StepFun/Step-5-Preview and tag as NEW. Its companion support PR, #44356, is still open. The cookbook page states plainly that "Step 5 Preview support has not shipped in an SGLang release yet, so a stock pip install sglang cannot serve it" — which is the single most important sentence on it.

Read those four together and you get a coherent picture: StepFun is standing up open serving for a model whose API has been live for nearly a month. Nothing here is a launch announcement, and nothing here is a benchmark. It is the plumbing that determines whether the model is something you can run.

What Step 5 Preview is, in the terms a serving engineer asks about first

StepFun's own English documentation is thin on architecture — it publishes the model ID, the 1M-token context window, text/image/video input, text-only output, a 64k-token maximum output, JSON Schema support, prompt caching and a reasoning_effort control with low, medium and high settings. It publishes no parameter count. The architecture figures below come from the two first-party serving documents published today, which had to read the checkpoint to write a loader, and they agree with each other.

• Size — 600B total parameters with 27B active per token, a sparse mixture-of-experts design. The vLLM recipe independently states 604B, "counted from the checkpoint tensor shapes," and the two numbers are the same number rounded differently.

• Depth and routing — 92 layers. Each MoE layer routes a token to 8 of 352 experts plus one shared expert. That expert count is the reason the vLLM PR carries a router-GEMM kernel change: it adds (4096, 352) as a supported (hidden_dim, num_experts) pair, alongside the existing (3072, 256) and (6144, 128).

• Attention — a four-layer repeating pattern: three sliding-window layers attending to the most recent 512 tokens, then one sparse-attention layer. The sparse layers split KV history into 8-token regions, score a summary of each region with a lightweight indexer, and select up to 512 historical regions to attend to at token level. SGLang calls it region-sparse attention; the practical consequence is that a 1M-token context stays affordable on memory.

• Speculative decoding — three MTP layers. SGLang's "three-layer MTP" and the vLLM recipe's --spec-method mtp --spec-tokens 3 are the same claim from two directions.

• Precision — routed-expert weights are FP8 (E4M3, 128×128 block scales) except in the last three MoE layers, which stay BF16. A separate NVFP4 checkpoint, stepfun-ai/Step-5-Preview-NVFP4, is referenced by the vLLM recipe as a variant. Both stacks leave the mixed-precision layer split intact rather than upcasting it.

• Reasoning behaviour — the model always thinks before answering. reasoning_effort sets the depth and defaults to high when unset, which is the setting most likely to surprise you on a latency budget. StepFun's recommended sampling is temperature=1.0, top_p=0.95, top_k=0, and SGLang's cookbook flags those as informational rather than something to hardcode.

The hardware floor is the part that decides whether this matters to you

This is where the vLLM recipe earns its keep, because nobody else has published hard minimums.

• FP8 floor — 773 GB of VRAM, with a minimum of 4 GPUs and a default hardware target of B300. Verified configurations are listed for H200, B300 and GB300. On four B300s the recipe offers TP4, TEP4 or DEP4 layouts; the FP8 weights do not fit four H200s, so the Hopper path is eight GPUs.

• NVFP4 floor — 473 GB of VRAM for the quantised checkpoint, 300 GB lower than FP8. The trade is context: the NVFP4 variant runs with --max-model-len 320000, so you give up roughly 724,000 tokens of the advertised 1,048,576 to fit in the smaller footprint.

• Expert parallelism — on Blackwell, TEP and DEP layouts route through MegaMoE with --moe-backend deep_gemm_mega_moe; the pure tensor-parallel layout deliberately does not, because MegaMoE requires expert parallelism to have anything to split. On H200 the recipe falls back to a Triton MoE for every expert layer, including the three BF16 ones, because the TRT-LLM MoE kernels are Blackwell-only.

• Startup cost — the SGLang cookbook puts the first launch at about 15 minutes of FlashInfer compilation on B300, dropping to about 6 minutes on subsequent starts, and tells you to persist FLASHINFER_WORKSPACE_BASE and ~/.cache/sglang across restarts. If you are autoscaling from zero, that is a cold-start floor, not a rounding error.

• What is not supported yet — the vLLM recipe marks both Simple and LMCache KV offload as unsupported for this model. SGLang's HiCache path is documented as available in the cookbook, which is a real capability difference between the two stacks as they stand today.

The practical reading: Step 5 Preview is a four-to-eight accelerator deployment, not a single-node hobby project. If that sentence ends your interest, the API section below is the rest of the article.

Calling it, and what each token costs

StepFun's own pricing page lists step-5-preview at $1.00 per million input tokens on a cache miss, $0.10 on a cache hit, and $2.70 per million output tokens — figures from the vendor, not from a reseller. Two details change the arithmetic. First, cache-miss input pricing includes writing new content to the cache, so a first pass over a long document is priced once and cheap on the second pass. Second, output tokens include the model's reasoning as well as its answer, which matters more here than on a model that answers directly.

Artificial Analysis, the one neutral evaluator with a published run, scores Step 5 Preview at an Intelligence Index of 43.7 — displayed as 44 — with a median output speed of 88.5 tokens per second and a time to first answer token of 24.7 seconds, against roughly 2,000 reasoning tokens per response. The speed figure is respectable and the answer-token latency is not: a model that always thinks first is a poor fit for interactive autocomplete and a fine fit for batch and agentic work, which is precisely how StepFun positions it. Those are independent numbers on a third-party harness, and they are the only ones available — StepFun publishes no benchmark table for this model at all.

Where the story is still soft, and it is not all soft in the same way

Three things are worth flagging rather than smoothing, because each one has a different shelf life.

The first is whether the vendor checkpoint is actually downloadable. Both first-party serving documents reference stepfun-ai/Step-5-Preview and stepfun-ai/Step-5-Preview-NVFP4 as the canonical checkpoints, and SGLang's cookbook links a LICENSE.txt under the StepFun Model License 1.0. What this page cannot do is confirm from Hugging Face that those repositories are publicly readable today: the Hugging Face API returns the identical 401 Invalid username or password for stepfun-ai/Step-5-Preview as it does for a repository id that has never existed, so the response cannot distinguish a gated model from an absent one. The recipe's own guidance — "use hf auth login if the checkpoint requires authentication" — is the only statement about access, and it is conditional. Treat the download as referenced by the people who wrote the loaders, and confirm access yourself before you plan a launch around it.

The second is the release date, which three sources give three ways: StepFun's announcement materials put the launch at September 20, 2026; Artificial Analysis dates the model September 18, 2026 in its own metadata and displays "Released September 2026"; and the model-metadata aggregators that most tooling reads carry September 16, 2026. None of the three explains the gap. This article uses the vendor's September 20 where a date is needed, and it is worth knowing which number the tool you are reading got.

The third is maturity, and here the two stacks are not equivalent. SGLang's cookbook is merged and merged into main, but SGLang's own text says support has not shipped in a release, so the Python path is a preview Docker image and nothing else. vLLM's recipe is open, its model PR is an unmerged draft that GitHub currently reports as having merge conflicts and needing a rebase, and its own reviewers have already raised three open items against the recipe — an FP8 hardware gate, a default B300 profile that emits eight-GPU commands against a four-GPU guide, and an NVFP4 context cap that is not encoded in the variant arguments. None of that is disqualifying. All of it argues for treating a first self-hosted deployment this week as an evaluation, not a migration.

What to do with this, this week

If you already have four B300s or eight H200s and a workload that looks like long-document analysis, multi-turn tool calling or overnight agent runs, the vLLM recipe is close to a runbook: pull vllm/vllm-openai:step-5-preview, copy a TP4 or TEP4 command, and check whether --max-model-len 1048576 survives contact with your actual KV budget before you promise anyone a million tokens. If you have Hopper and not Blackwell, start with the NVFP4 checkpoint and accept the 320k context, or budget eight H200s. Either way, set reasoning_effort explicitly on every request rather than inheriting the high default, and measure the 24.7-second time-to-answer-token on your own traffic before you conclude anything from the 88.5 tokens per second figure.

If you do not have that hardware, the API is not a consolation prize — it is the same weights, billed per token, at $1.00 in and $2.70 out with a 90% cache discount on repeat context. And here is the part of this page where the honest answer about us belongs: Step 5 Preview is not on OrcaRouter's live model list as of today. Our catalogue is 207 models across 13 providers and it currently contains no StepFun model at all, so for this particular model the vendor's own API is the only callable path, and nothing on this page should be read as a claim that we serve it. What we do route is the surrounding decision: when you are picking which frontier model carries a production workload, the argument for reaching it through one endpoint is that provider list price is passed through at 0% markup, so a vendor's price change is your price the same day, and that automatic failover across providers means a single model's bad afternoon is not your outage. That is a statement about the models we do carry, and it is deliberately not a statement about this one.

What to watch next is narrow and checkable. Whether stepfun-ai/Step-5-Preview resolves to a readable repository with a licence file you can point a legal team at; whether vLLM's #61727 clears draft and its merge conflict; whether SGLang's support PR #44356 lands in a tagged release, which is the point at which the recipe stops being a preview image and becomes something your CI can pin. Any one of those three turns this week's "you can probably run it" into a plain yes, and until then the only claim worth making is the one the evidence supports: on October 15, 2026, two serving stacks published their own paths to a model that has been callable since September 20, 2026.

Choosing to run a 600-billion-parameter model on your own hardware is also the kind of decision where the track record matters: documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.