StepFun's Step 5 Preview Launch Night: What Five Builders Said About a Model That Shipped Six Weeks Earlier

It rained on the way in, which is the only part of the evening in San Francisco on 22 October 2026 that StepFun could not control. The company rented a room, co-hosted with Together AI, and filled it — around 277 people had signed up on the public event page before the doors opened — for what the invitation called the Step 5 Preview Launch Night and what the calendar called a Thursday. The model being launched had an API since 20 September and downloadable weights since 15 October. Neither of those dates moved for the party. What the party did was put the people who build on top of models in one room with the people who trained this one, and let them talk about the parts of the work that a benchmark table never captures.

The panel was small and unusually specific: Hongyu Zhou, a pretraining researcher at StepFun; Dan Fu, who runs kernels at Together AI; Mar Hershenson of Pear VC; Saoud Rizwan of Cline; and Joey Wang of Harvey. Between them they covered the four things that decide whether a model gets used — how it was trained, how it is served, whether anyone will fund the thing built on it, and whether an agent loop can rely on it not to fall over at step forty.

First, the date problem, because it explains the whole evening

Step 5 Preview is not new. StepFun opened the commercial API on 20 September 2026, and the launch ran in preview form from the start — step-5-preview is the model ID, not a modifier somebody added later. Artificial Analysis dates the release to "September 2026" without a day, and its older model page put the date at 18 September; the discrepancy with StepFun's own date has never been explained by either party, and this article uses StepFun's 20 September with that caveat attached.

The date that mattered on 22 October was 15 October. That is when the BF16 and FP8 checkpoints went up on Hugging Face and ModelScope, alongside a technical blog describing a 92-layer "narrow-deep" stack and a sparse GQA attention pattern that interleaves gated sliding-window layers with sparse full-attention layers. The event was, in the plainest reading, a launch night for a weights release: the evening the model stopped being something you call and became something you can also run.

That reading is worth stating because the alternative one — that this was a new model — was in the air all evening, and it is wrong. Anyone who came away thinking StepFun shipped a model in late October will be six weeks late to a decision they could already have made.

What the room was actually shown

The technical content was lighter than the marketing, which is normal for this format. What StepFun put on screen was the model's shape rather than a fresh benchmark run: a sparse mixture of experts with 600 billion total parameters and 27 billion active per token, 8 of 352 experts routed plus one shared expert, three multi-token-prediction layers for speculative decoding, and a 1M-token context window. It accepts text, images and video and returns text only. The output ceiling in StepFun's own documentation is 64k tokens.

Zhou's segment was the one that got closest to a real claim about why the architecture looks like this: a deep stack of narrow layers, on the argument that depth buys implicit multi-hop reasoning that a shallower, wider model has to spend tokens doing out loud. Whether that argument is right is not something a launch night can settle. It is, however, the kind of claim that a downloadable checkpoint lets other people test, which is exactly what was not possible before 15 October.

Tool calling was the topic. Benchmarks were not.

The panel's most useful stretch was about reliability in agent loops, and it came from the two people in the room who spend their days inside them rather than training models. Rizwan's product, Cline, is an editor agent that has to keep working when a model returns malformed JSON at the eleventh step of a twenty-step task. Wang's employer, Harvey, runs similar loops for legal work where a wrong tool call is not a bad demo but a liability event.

What both described is a failure mode that does not show up in a benchmark average: the model that is excellent for ten turns and then silently changes the shape of its arguments. StepFun's published evidence for that dimension is its own, and it is worth labelling as such. On Terminal-Bench 4.0 the vendor reports 33.3 against 12.6 for Kimi K3; on DeepSWE v1.1, 67.7 against 67.5. Those are vendor-run numbers on the vendor's choice of harness, and the gap on terminal work in particular is large enough that it deserves a second opinion before anyone plans a migration around it. The independent evaluator has not published a harness-matched head-to-head for this model, and the index it does publish has since been rebuilt — see below.

Inference cost, discussed as a serving problem rather than a price list

Fu's segment was the one that will matter most to anyone who intends to self-host, and it was specific in a way that launch talks usually are not. The FP8 checkpoint needs roughly 773 GB of VRAM and a minimum of four GPUs, with B300 as the default target; the Hopper path needs eight H200s because the model does not fit on four. The NVFP4 variant drops the footprint to about 473 GB and caps usable context at 320k tokens, which is a real trade rather than a rounding error. vLLM and SGLang both had recipes out on 15 October, and the SGLang path pays a FlashInfer compilation cost of around fifteen minutes on a cold start.

Against that, the rented alternative is simple: $1.00 per million input tokens and $2.70 per million output tokens on StepFun's dollar card, with cached input billed at a 90% discount. StepFun's renminbi card reads ¥7 input, ¥20 output and ¥0.70 for cache hits — the cache-hit figure having doubled from ¥0.35 at 00:30 Beijing time on 14 October, eight days before the party, with a grandfather clause for earlier pay-as-you-go callers. Nobody on stage mentioned that change. It is the single most consequential pricing fact about this model this month, and it is in the vendor's own documentation rather than in the deck.

Artificial Analysis puts the blended cost at $0.54 per million tokens at a seven-to-two-to-one cache-hit ratio, and the cost of a full Intelligence Index run at $2.35.

The capital question, which is where the evening got honest

Hershenson's contribution was the least technical and the most useful: a reminder that the reason to care about a cheap, open, long-context model is that it changes what a seed-stage company can attempt. A startup that could not have afforded a frontier model at every step of an agent loop can now afford to run the loop more than once. Wang's version of the same point, from inside an enterprise buyer, was that the procurement question is not "is this the best model" but "what happens to our costs and our latency if it is twice as good as we expected and we route everything to it".

That is a routing question, and it is the one place where self-hosting economics and API economics stop being separate conversations. OrcaRouter's position on it is straightforward: the platform carries 214 models behind one key and one bill, with a routing configuration that lets you declare a primary and a fallback chain rather than hard-coding a model ID into an application, and automatic failover when a provider degrades. Step 5 Preview is not among those 214 — there is no StepFun entry in the live catalogue as of 29 October 2026, and nothing in this article should be read as a claim that we serve it. Anyone who wants to evaluate it today is choosing between StepFun's own API and self-hosting the checkpoints that landed on 15 October.

What actually changed after the model shipped

Three things, in order of how much they should affect a decision.

The first is the index. On 27 and 28 October Artificial Analysis replaced its Intelligence Index with version 5, weighting 60% of the composite on private or held-out test sets and rebuilding the nine constituent evaluations. It dropped three of the evaluations that StepFun's launch material leaned on — GDPval-AA, SciCode and AA-LCR — and added FrontierCode 2.0 and Terminal-Bench-Science. Step 5 Preview's score on the new index is 29. Its score on the old one was 44. Those are two different measurements of the same model on two different rubrics, not a model that got worse, and the 44 that circulated in September and October should not be subtracted from anything scored under v5. Kimi K3.1 Pro, for comparison, moved from 55 to 46 across the same rebuild — a nine-point drop on a model that had just launched.

The second is the licence. The weights are open in the sense that you can download them, and they are not open in the OSI sense. StepFun Model License 1.0 keeps an attribution condition on materially used outputs, requires hardware makers — phones, vehicles, wearables, robots — to negotiate separately, and requires a commercial licence for model-as-a-service operators above $20 million in annual revenue. If you are a two-person team, none of that bites. If you are planning to resell inference on top of it, it bites immediately, and it is worth reading before the first GPU hour rather than after.

The third is the price change already described, which happened before the event and did not come up at it.

What a builder should take from the evening

The useful signal from 22 October is not that StepFun has a new model. It is that a company with a flagship it released in September spent an evening in October talking to agent builders about tool-call reliability and kernel-level serving costs, in a room where the people who would suffer from both were sitting in the front rows. That is a company behaving as though the next six months of adoption depend on operability rather than on another benchmark row.

The concrete version of that: if you were waiting for weights before you would take Step 5 Preview seriously, the wait ended on 15 October. If you were waiting for an independent score you can compare against everything else, the number now exists, it is 29, and it is measured under a rubric that has just been rebuilt underneath everyone. Both of those are reasons to form a view now rather than after the next launch night.