Step 5’s Open Weights Are Out — How to Download, Run and Decide If You Should

Last updated: October 15, 2026

StepFun’s Step 5 open weights are live — a day that started with a missing release and ended with one. The Shanghai lab promised the BF16 weights for its 600-billion-parameter flagship on 15 October, and by morning the download wasn’t there. By afternoon IST it was: the official Hugging Face repo now carries the full weights, the licence text, and the serving code. This guide picks up where our Oct 15 preview checklist left off — what’s actually in the release, the licence terms you must read before commercial use, the hardware it honestly takes, and the API-vs-self-hosting maths.

Quick answer: Step 5's open weights landed 15 Oct 2026 after a morning delay. What's in the download, the StepFun licence, hardware needs, and API ($1/M) vs self-hosting.

What happened on 15 October — the honest timeline

The weights arrived late, and the delay is itself the story. StepFun announced Step 5 Preview on 20 September and committed to publishing the BF16 open weights on 15 October 2026. When Forkast checked that morning at 01:55 UTC, the expected repository returned a 404, no licence was published, and the lab’s official X account had posted nothing about a release (Forkast, 15 Oct 2026). Unendorsed third-party mirrors had already appeared — none of them carrying the lab’s blessing.

Then the release landed. A desk check of the official repository (stepfun-ai/Step-5-Preview-BF16) on the afternoon of 15 October IST found it fully populated: 48 files including the model-*.safetensors BF16 shards, LICENSE.txt, config.json, the chat template, and a complete inference/ directory with conversion and generation code. The repo was created back on 20 September — the announcement day — but the files only landed today (last-modified timestamp 01:31 UTC, roughly 07:01 IST; downloads were already in triple digits by afternoon). The morning’s 404 was most likely upload propagation or a visibility toggle, not a cancellation.

Treat that sequence as the lesson in miniature: a lab’s announced date is a promise, not a download. Plan infrastructure around the artifact, not the announcement.

What’s actually in the release

The release is complete: weights, licence, config, and serving code in one place. Beyond the safetensors shards, the repo ships the model configuration, the chat template, a vision encoder, and inference code covering vLLM, Transformers and SGLang deployments — including the reasoning and tool-call parsers the model needs.

The specs, confirmed on the official model card: a sparse mixture-of-experts model with 600 billion total parameters and about 27 billion active per token, up to 1 million tokens of context, text/image/video input and text output, with low/medium/high reasoning effort settings. The card’s evaluation notes cite the model’s Artificial Analysis scores as of 20 September 2026 — the one fully independent number remains 44 on the Artificial Analysis Intelligence Index, ranked 27th of 653 models tracked, at a weighted cost of $0.71 per completed task (via GenZTech’s October analysis).

Skip the third-party mirrors now. SHSLab and similar copies were the only option during the morning gap; they are unendorsed by the lab, and the official repo is live. There is no reason left to trust a stranger’s upload of a 1.2 TB file.

The licence is not Apache 2.0 — read this first

This is the most important new fact of the release: the weights ship under “StepFun Model License 1.0”, a custom licence — not the Apache 2.0 of StepFun’s earlier Step 3.5 Flash. The full licence text was read in the repo so you don’t have to guess. The plain-English version:

For a hobbyist, researcher or small team running the model internally: this licence is generous. For a hardware maker or an API reseller at scale: talk to StepFun first. And for everyone planning commercial use — read LICENSE.txt yourself; the summary above is not legal advice. The licence question our preview checklist flagged as “TBD” now has an answer, and the answer is “custom, not Apache”.

Can your hardware run it? The honest table

Open weights do not make a 600B model cheap to run — they make it yours to run. The official model card publishes exact deployment configurations, and they are datacenter hardware:

FrameworkPrecisionHardwareParallelism
vLLMFP8One B300 node (4 GPUs)TP4 / TEP4 / DEP4
vLLMFP88x H200TP8 / TEP8 / DEP8
vLLMBF16One B300 node (8 GPUs)TP8 / TEP8 / DEP8
vLLMBF16Two H200 nodes (16 GPUs)TP16 / TEP16 / DEP16
TransformersBF16Two H200 nodes (16 GPUs)TP16 / EP16
SGLangFP8One B300 node (4 GPUs) or 8x H200TP4/EP4 or TP8/EP8

The full-precision weights need roughly 1.2 TB of memory before KV cache (MarkTechPost) — that is 8 to 16 datacenter GPUs, not a gaming PC. Community quantised builds (GGUF, exl2) will shrink this dramatically, but they do not exist yet and depend on the licence permitting them — which, on the reading above, it does for non-hardware, non-MaaS use. Anyone telling you a home rig runs the full weights today is selling you something.

How to download the weights

The download itself is the easy part — the disk space is not. Numbered steps:

  1. Free ~1.5 TB of fast storage. The BF16 shards are about 1.2 TB; you want headroom for the download cache and conversion outputs.
  2. Install the Hugging Face CLI: pip install -U "huggingface_hub[cli]".
  3. Pull the official repo only: hf download stepfun-ai/Step-5-Preview-BF16 --local-dir ./step5-bf16. Double-check the org is stepfun-ai — the official account — before a terabyte starts moving.
  4. Verify the file list. You should see model-*.safetensors shards, config.json, LICENSE.txt, chat_template.jinja, and the inference/ directory. If a file is missing or zero-byte, re-run the download rather than patching it from a mirror.
  5. Read the model card’s deployment section before serving. The repo ships vLLM, Transformers and SGLang configs with the exact parallelism flags — start from those rather than inventing your own.

API vs self-hosting — the honest maths

For most teams, the API is still the right answer. The numbers, from the published listings:

RouteInputOutputCached inputHardware needed
StepFun API / OpenRouter$1.00 / M tokens$2.70 / M tokens$0.05 / M tokensNone
Self-hosted (your GPUs)$0 per token$0 per tokenn/a8-16 datacenter GPUs, power, engineering

Self-hosting wins when you have sustained, high-volume traffic and the GPUs are already paid for — or when you need to fine-tune, inspect, or keep data fully on-premises. The API wins on everything else: zero capex, 1M-token context out of the box, and cache pricing that makes long agent loops cheap. Run your real workload through the API first and log the cost — that measurement is the baseline your self-hosting decision has to beat. (Step Plan’s monthly credits expire at month end and don’t roll forward, per Omidsaffari’s pricing analysis — don’t pre-buy a big tier speculatively.)

The lesson the morning delay taught

Split every model release into three milestones: announced capability, usable hosted access, and downloadable artifact. The first two were true for Step 5 since September; only today did the third arrive — and for several hours this morning, even the announcement’s date proved soft. Teams that scheduled GPU procurement or migration work around “Oct 15” bought themselves a morning of nothing.

The durable rule: never commit infrastructure, hiring, or launch dates to a vendor-promise date. Commit them to the artifact. Step 5’s weights are an artifact now — and this guide is the playbook for using them.

FAQ

Is Step 5 Preview really open source?

Not under an OSI-approved licence. It ships under the custom “StepFun Model License 1.0”: free to use, modify and redistribute, with attribution required at large scale and a separate commercial licence needed for hardware-device features or large model-as-a-service businesses. Open weights, yes; open source in the strict sense, no.

How big is the download?

Roughly 1.2 TB in BF16 before KV cache. Budget ~1.5 TB of disk for the download plus working space.

Can I run it on a single GPU?

No — not at full precision. The official serving configs start at 4x B300 or 8x H200 (FP8) and 8x B300 or 16x H200 (BF16). Wait for community quantised builds if you have modest hardware.

What happened to the morning delay?

StepFun committed to 15 October; at 01:55 UTC the official repo still returned a 404 with no licence and no announcement (Forkast). The files landed around 07:01 IST the same day. A late release, not a cancelled one.

Is the API still worth it now that weights are out?

Often yes. It remains the zero-hardware route with 1M-token context and $0.05/M cached reads. Self-hosting pays off at sustained high volume or when you need fine-tuning and data control.

Sources and methodology

Claims rest on the sources above plus a direct desk check of the official Hugging Face repository and its licence text on 15 October 2026. Vendor-published benchmark claims are labelled as such; the Artificial Analysis index is an independent measurement. Desk-researched, not hands-on tested — no GPU cluster was harmed. Verification date: 15 October 2026.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!