Step 5 Preview: Official Weights or API Access?
Verify the official Step 5 Preview checkpoint, understand 28 files versus 26 indexed shards, and choose between hosted access and deployment evaluation.
Ofox 
Step 5 Preview BF16 now has weight files listed in the official StepFun Hugging Face repository. On October 15, 2026, we retrieved its metadata, model card, configuration, license, and weight index. The repository lists 28 safetensors files, while the main weight index references 26 distinct files. Those are different counts, not conflicting reports.
This is a release-artifact check. We did not download the full checkpoint, hash every weight file, run inference, or benchmark the model. If you want to use the model today, separate three questions: are the official artifacts listed, can your infrastructure run the chosen format, and does a hosted provider give your account access?
What we checked, pinned to one version
The primary source is stepfun-ai/Step-5-Preview-BF16, pinned to commit:
7734b7ffac118d70d8398bfa3eb0fab007eb45b1The repository metadata reported an update at 2026-10-15T01:31:42Z. A cached search snippet saying that weights are “coming October 15” can therefore lag the repository. The live branch may change again; the fixed revision makes this check reproducible.
| Artifact | Observed evidence | What it does not prove |
|---|---|---|
| Repository metadata | 28 listed .safetensors files | Successful transfer or integrity of every file |
| Main weight index | 26 unique referenced files, all found in the listing | End-to-end loading or inference |
| Configuration | Step5ForConditionalGeneration, BF16 text dtype | Compatibility with an arbitrary installed runtime |
| Model card and inference notes | Downloaded and read at the pinned revision | Our own GPU validation |
| License | StepFun Model License 1.0 text available | Unconditional permission for every commercial use |
The index’s metadata.total_size is 1,208,684,168,984 bytes, about 1.209 TB in decimal units or 1.099 TiB in binary units. This describes the indexed tensor size. It is not a measured full download, total disk requirement, or minimum GPU memory specification. Extra files, conversions, temporary copies, runtime buffers, and KV cache need separate planning.
Verify the listing without downloading a terabyte
Download the metadata-check script. It uses Python’s standard library and fetches only the small repository metadata and index, never the weight files.
python3 step_metadata.pyThe script prints the pinned commit, total safetensors file count, distinct indexed file count, indexed tensor bytes, and any referenced filenames missing from the repository listing. For the revision above, expect 28 files, 26 indexed files, 1208684168984 indexed bytes, and an empty missing-file list.
A different result is a reason to inspect the metadata, not a reason to silently substitute the latest revision. HTTP or JSON errors should stop the check. To inspect the underlying evidence yourself, open the fixed model card and fixed weight index.
Passing this script means the index and listing agree. It does not mean the complete model has been downloaded, that all remote objects were byte-verified, or that inference will succeed. Save the output with the time of your check before evaluating a different checkpoint revision.
Choose the precision before choosing hardware
The official model card links the FP8 repository and the BF16 repository examined here. Follow the link for the precision you actually intend to deploy; do not mix a tokenizer, configuration, converted shards, and weights from unrelated revisions.
The model card describes 600B total parameters with 27B active per token. Active parameters describe computation, not permission to load only 27B parameters. Treating “27B active” as the whole storage or memory footprint would produce an unusable deployment plan.
Official starting layouts include:
| Checkpoint | Example layout in the model card |
|---|---|
| FP8 | 4 × B300 or 8 × H200 |
| BF16 | 8 × B300 or two nodes totaling 16 × H200 |
These are vendor example layouts, not minimum requirements measured by Ofox. They also do not guarantee the advertised maximum context will fit. The configuration’s 1,048,576 position limit is a model capability setting; available memory, batch size, and KV cache determine what a particular deployment can serve.
A BF16 checkpoint is required for the BF16 path. Setting --dtype bfloat16 while loading FP8 files does not convert those files into the official BF16 checkpoint.
Read the runtime caveats before provisioning
The model card references a Step 5-specific vLLM image, vllm/vllm-openai:step-5-preview, and framework support for the model and relevant parsing/speculative features. A generic “latest” installation is not evidence that those components are present.
The pinned inference README lists compatible dependencies including torch>=2.10,<2.14, transformers>=4.57.6,<6, safetensors>=0.7,<1, and triton>=3.6,<3.8. CUDA PyTorch must also match the host drivers. Check these constraints against the exact runtime you deploy instead of assuming that an older environment is suitable.
The standalone conversion path shards the checkpoint while preserving its dtype; it is not BF16-to-FP8 quantization. Tensor and expert parallelism share a group in the documented path, so TP8 and EP8 do not mean 64 GPUs. Converted weights, configuration, and tokenizer must remain matched.
Most importantly, the model card says that the complete BF16 convert.py → generate.py --tp-dir file-sharded flow has not been validated end to end. The inference notes separately describe official two-node full-model validation with streamed BF16 weights. That is meaningful evidence for that path, but it does not remove the first caveat. A reported allocation for a specific test is not a universal memory requirement for all contexts and runtimes.
If you need images or video, also inspect processor and prepared-input requirements. A text-generation smoke test is not validation of visual preprocessing, OCR quality, or video understanding.
Weights, self-hosted endpoints, and hosted API access
| Your task | Sensible next route | Acceptance evidence |
|---|---|---|
| Inspect release status or license | Official repository and small metadata files | Pinned revision and matching artifact list |
| Evaluate inference infrastructure | Chosen checkpoint plus its matching runtime instructions | Load success, a known task, resource measurements, and error logs |
| Use a managed service without hosting weights | Provider’s current model and account documentation | Exact endpoint, model ID, account permission, and reconciled usage |
The official hosted Step 5 Preview model documentation is the place to check the current cloud route. This article does not supply an unverified cloud request or claim that an account can call it. The step5 served-model name in a self-hosting example is a local server alias, not proof of the official cloud API identifier.
For a hosted evaluation, record the provider, endpoint, exact model identifier, region/account eligibility, price source, and data policy before making an authorized request. Then verify HTTP status, returned model information, expected output, usage, and billing. A repository download or successful request to a different Step model does not complete those checks. We have not verified an Ofox Step 5 route in this article.
For self-hosting, begin with the documented small smoke test after the full artifact and runtime checks. Keep output acceptance separate from launch success: a process listening on a port does not establish valid model output. On failure, preserve the exact revision, dtype, runtime build, parallel layout, and first loading error before changing multiple variables.
Check the actual license
Read the pinned LICENSE.txt. The file is StepFun Model License 1.0, not an unqualified Apache-2.0 license. Although the model card uses Apache-style wording, the license contains additional attribution and commercial conditions, including specified business circumstances requiring separate permission.
Do not translate “open weights” into “unconditional commercial use.” Keep the license revision with your deployment evaluation and have the relevant owner assess your intended use against the actual terms.
You have completed this article’s task when you can identify the official checkpoint and revision, reproduce the listing/index distinction, select the matching precision and runtime path, and state which deployment or account checks remain. Performance rankings and real deployment results require further evidence; artifact availability alone cannot supply them.
