Step 5 Preview's Weights Were Never Missing: What the NVFP4 Checkpoint Changes

S​tep 5 Preview has been downloadable since the day StepFun released it on September 20, 2026 — and a large share of what has been written about the model since then says otherwise. The FP8 repository stepfun-ai/S​tep-5-Preview was created at 03:15 UTC that morning and its first shard was committed 36 minutes later; the BF16 repository followed at 03:52 and its shards landed at 04:18. Both are StepFun's own, both are ungated, and together they carry 1.86 terabytes of weights. What changed in the middle of October is narrower and more useful than a re-release: StepFun published a third checkpoint in N​VIDIA's NVFP4 format, and on October 15, 2026 N​VIDIA publicly endorsed it — calling this "their largest open weights release," saying the model "is built to handle agent tasks lasting up to 24 hours," and putting a number on it: "nearly 8K tokens/sec/GPU on N​VIDIA GB300 using NVFP4."

That is the news. The correction is the reason this piece leads with a date instead of an announcement. Coverage written on September 20 ran against a repository that was minutes old and briefly empty, and the "weights promised for a later date" framing stuck even after the weights arrived the same morning. A sibling article on this site repeated it as recently as October 13. It is not true now and it was not true then, and it matters because the open-weights question was the single largest unresolved item in every write-up of this model.

What actually shipped on October 14 and 15

The new artifact is stepfun-ai/S​tep-5-Preview-NVFP4. The repository was created at 02:59 UTC on October 14, 2026 and the weight upload is a commit at 07:10 UTC the same day by a contributor whose account, huangyu-nv, also rewrote the README an hour later to align it with the FP8 and BF16 repositories. It holds 394.18 GB across 95 safetensors files arranged as 92 per-layer files plus two extras and one vision tower file — a different sharding scheme from the FP8 and BF16 repos, which both ship 24 text shards plus a multi-token-prediction head and a ViT.

What is in it is worth stating precisely, because "NVFP4" is being used loosely in circulation. The quantization block in config.json reads quant_method: modelopt, quant_algo: NVFP4, group_size: 16, and kv_cache_quant_algo: null. The ignore list is long and specific: 92 self-attention blocks, 88 MoE gates, 88 shared experts, the embedding table, the final norm and the language-model head, plus the vision tower and its projector. Only three entries in the whole list touch moe.experts. In plain terms, attention and the routing machinery stay in higher precision and the expert feed-forward weights — the bulk of the 600 billion parameters — are the part that got quantized. The KV cache is untouched, which is the detail that determines whether your long-context serving configuration survives the switch.

Hugging Face's own parameter accounting on that repository reads 320.6 billion, split as 282.4 billion U8, 38.2 billion BF16 and 2.4 million F32, and the file page prints "321B params" as the model size. StepFun's card, on the same page, says 600B total and 27B active. Both are correct and they are counting different things: the 321B figure is a storage count of quantized elements, not a parameter count in the sense the architecture description uses. If you are sizing memory, use the storage number. If you are describing the model, use the card.

Three checkpoints, three hardware bills

The reason the NVFP4 release matters more than a fourth file in a folder is that it changes the smallest machine you can run this model on. StepFun's own deployment table in the model card gives the shapes for the two older checkpoints and the repositories give the sizes; here they are together, with the NVFP4 row added.

• Precision — FP8 650.32 GB vs BF16 1,214.9 GB vs NVFP4 394.18 GB on disk

• Smallest published vLLM configuration — FP8 one B300 node with 4 GPUs, or one H200 node with 8 GPUs vs BF16 one B300 node with 8 GPUs, or two H200 nodes with 16 GPUs vs NVFP4 not yet in StepFun's table

• Parallelism the card recommends — FP8 TP4, TEP4 or DEP4 on B300 vs BF16 TP8, TEP8 or DEP8 vs NVFP4 not stated

• Attention and router weights — FP8 quantized to eight bits vs BF16 full precision vs NVFP4 left unquantized by the ModelOpt ignore list

• KV cache — unquantized in all three, since the NVFP4 config sets kv_cache_quant_algo to null

• Serving stack — all three through vLLM or SGLang with an O​penAI-compatible endpoint, or offline through the standalone Transformers path

The gap in the last two rows is the honest caveat on the new checkpoint. N​VIDIA's post says the model runs at nearly 8K tokens per second per GPU on GB300 using NVFP4, but StepFun's deployment table has not been updated with an NVFP4 row, so there is no vendor-published parallelism recipe for it yet. If you are planning a deployment this week, the FP8 path is the one with a documented configuration and the NVFP4 path is the one with a headline number attached to somebody else's hardware.

The 8K figure is a vendor number, and only one vendor has said it

N​VIDIA's October 15 post is the entire first-party record on this. There is no N​VIDIA developer blog entry, no StepFun announcement page, and no third-party reproduction of the throughput claim that can be checked from outside. "Nearly 8K tokens/sec/GPU" is also a statement that means very little without the batch size, the sequence length and the concurrency it was measured at — the same model configuration can differ by an order of magnitude across those three axes, and serving frameworks for mixture-of-experts models are specifically sensitive to how many experts are active per token. Treat it as N​VIDIA reporting on N​VIDIA hardware, which is exactly what it says it is, and not as an independently verified throughput figure.

The 24-hour agent claim has the same status. It is a capability description rather than a measurement, and StepFun's own card describes long-horizon execution in the same terms without attaching a duration to it. The vendor's published evaluation notes for this model do list the harness and the settings — temperature 1.0, top-p 0.95, top-k 0, the High reasoning setting, the SWE-agent harness for DeepSWE v1.1 — which is more than many release cards offer, but the score table itself is a picture and StepFun publishes no numeric benchmark table in its documentation. The only neutral evaluator with a live score for this model is Artificial Analysis, where S​tep 5 Preview sits at 44 on the Intelligence Index and originally cost $1.03 per index task to run.

The deployment recipe landed with the checkpoint, and it is the part you will use

Reading the repository histories, the weight files and the instructions did not arrive together. Both older repositories were rewritten across October 13 and 14 — seven README commits on the FP8 repo between October 13 and 15 — and what those commits added is a full serving guide: a published container image at vllm/vllm-openai:step-5-preview, worked commands for tensor, expert and combined parallelism, the flag set this model needs, and a validated SGLang path. The flags are where the model-specific work is visible: --max-model-len 1048576, a speculative decoding head configured as --spec-method mtp --spec-tokens 3, and dedicated --reasoning-parser step5 and --tool-call-parser step5 parsers so that the model's reasoning blocks and tool calls survive the round trip through an O​penAI-shaped API. On Blackwell, the TEP and DEP configurations switch the MoE backend to deep_gemm_mega_moe; the plain TP4 configuration cannot, because that backend requires expert parallelism.

That is framework support landing, and for anyone deciding whether to build on this model it is worth more than the throughput claim. A 650 GB download with an unvalidated launch command is a research project. A 650 GB download with a pinned container image, four tested parallel configurations per precision and a parser that understands the model's own output format is a deployment.

If you do not want to own any of that, the arithmetic on this model is unusual. S​tep 5 Preview runs on StepFun's own platform at $1.00 per million input tokens, $0.10 per million cache-hit input tokens and $2.70 per million output tokens, so a month at 200 million input tokens and 200 million output tokens bills $200 and $540 respectively — against roughly the price of a four-GPU B300 node per month for the hardware that serves it. Self-hosting here is a control and data-residency decision, not a cost decision, and the weights being open is what makes it a decision you are allowed to make at all. It is worth saying plainly that S​tep 5 Preview is not on OrcaRouter's live model list, so the place to call it is the vendor's own API; what one API for more than 200 models does buy you in this situation is the other half of the question — the cost of the model you would move the workload to if a self-hosted deployment does not work out, at 0% markup on every provider's list price.

The licence, read literally

"Open weights" is doing work in N​VIDIA's phrasing that the licence does not fully support, and this is the part of the story most likely to be misread. StepFun Model License 1.0 is a perpetual, worldwide, royalty-free, non-exclusive grant to use, copy, distribute, sublicense, modify and create derivative works from the weights, with a matching patent grant. It is not an OSI-approved open-source licence, and there are two conditions worth knowing before you build a business on it.

• Attribution — if a commercial product or service incorporating the model passes 100 million monthly active users or $20 million in revenue in any calendar month, you must prominently display "Powered by StepFun AI" in the product interface and in your user agreement

• Separate commercial licence required — if you are a hardware manufacturer or brand owner shipping the model's capabilities in devices, or if you run a model-as-a-service business and combined revenue exceeds $20 million over any twelve months

• What is explicitly not MaaS — embedding the model's capability in a fixed end-user feature, or merely relaying requests to another provider's model

• What is explicitly permitted without the extra agreement — internal use that does not expose the model or its capabilities to a third party, and use obtained through a StepFun-certified distribution channel within the scope of that certification

• Termination — material breach ends the grant, with a thirty-day cure window after written notice

So the practical read is that the weights are genuinely downloadable and genuinely usable, including commercially, up to the thresholds in Section 2. A startup fine-tuning this model for internal document work is inside the grant. A handset maker shipping it as a feature, or a hosted inference business past $20 million, is not. That is the same shape of restriction the L​lama and Q​wen licences use, and it is worth stating plainly rather than filing under "open."

Where this leaves the model

A 600-billion-parameter mixture-of-experts model with 27 billion active parameters, a million-token context, text, image and video input, a 64k maximum output in StepFun's documentation, an approximately $1.00 / $2.70 rate card, an index score of 44 from the one neutral evaluator that has scored it, and three downloadable checkpoints at three precisions. Released September 20, 2026. Weights open since the morning of release. Quantized for Blackwell on October 14, endorsed by the GPU vendor on October 15, at a throughput figure that no one outside N​VIDIA has reproduced.

The next thing worth watching is not another endpoint. It is the first independent measurement on the NVFP4 checkpoint — a throughput or quality comparison against the FP8 path from someone who is not selling the hardware. Quantization at four bits with the attention stack held back is a reasonable design, and the specific ask is whether the expert weights survive it. Until somebody publishes that, the honest position on the new checkpoint is the same one the September coverage should have taken on the old ones: the weights are real, the recipe is real, and the performance claim is the vendor's.

A result like this is also a reminder that model mistakes are already documented: the AI incidents and agent failures behind 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.