STEP-5-PREVIEW-DERISKED-NVFP4
Manually gated NVFP4 derivative of the Step-5-Preview multimodal sparse Mixture-of-Experts model.
STATUS: Pending release authorization and quality assessment
Model weights are staged behind manual gating. Access requests will remain pending while this status is in effect.
Why this repository exists
This repository stages a candidate NVFP4 checkpoint and communicates its review status. The candidate is intended for controlled, authorized research and evaluation after its licensing, quality, capability, and security reviews are complete. This page is not a release announcement and does not authorize deployment or redistribution.
Artifact
| Model family | Step-5-Preview |
| Architecture | Multimodal sparse Mixture-of-Experts |
| Reported parameter scale | 600B total, 27B active per token; derivative verification pending |
| Format | NVIDIA NVFP4 mixed precision |
| Target hardware | NVIDIA Blackwell; final support matrix pending qualification |
| Inputs | Upstream text, image, and video capabilities; derivative qualification pending |
| Output | Text |
| Reported context | Up to 1M tokens upstream; 512K text-serving allocation and API smoke validated on TP4 B300, long-prompt quality pending |
| Repository contents | Manually gated model weights and index, sanitized configuration, tokenizer assets, pristine chat template, deployment kit, model card, and account banner |
The candidate is a behaviorally modified derivative. The proprietary transformation process is not distributed. No model-modification recipe or private research artifacts are included.
Lineage
StepFun Step-5-Preview model family
-> behaviorally modified derivative
-> NVFP4 release candidate
The model family and foundational lineage are StepFun's Step-5-Preview. No third-party mirror or community re-upload is identified as an upstream parent of this artifact. StepFun has not yet published an official source repository and immutable checkpoint revision that can be pinned here. That information, along with tokenizer and multimodal processor lineage, will be added when the official StepFun release is available.
The reported family-level specifications in this card are informational and do not establish checkpoint provenance. Final lineage, attribution, required notices, and redistribution authorization will be re-audited before release authorization.
Artifact and card verification
| Item | SHA-256 or status |
|---|---|
ASSETS/BLACKFROST-AI-BANNER.png | 52a12b0caf63c1859da518738f78517a49a2616a8bc04daecc515f8793e30f09 |
| Hugging Face metadata validation | Passed |
| Protected-information and repository-scope audit | Passed |
| Official StepFun source repository and revision | Pending official publication |
| Weight-index structural audit | Passed locally; final Hub verification pending |
Verification and evaluation status
| Review area | Status |
|---|---|
| Repository scope and protected-information audit | Complete for the staged weight publication |
| Model artifact publication | Staged behind manual gating; final Hub verification pending |
| Quality and capability assessment | Pending |
| Safety and behavioral assessment | Pending |
| Long-context and multimodal qualification | 512K text allocation/API smoke passed; long-prompt quality and multimodal validation pending |
| Release authorization | Pending |
No benchmark, throughput, refusal-rate, safety, or quality claim is made for the derivative at this stage. Upstream results must not be treated as measured results for this candidate.
Prompt, template, sampling, and deployment
The repository includes the pristine, non-baked chat template. It does not include a private system prompt or a model-modification recipe. No sampling recommendation is claimed while evaluation remains pending.
See DEPLOYMENT/README.md for the pinned SGLang
deployment procedure, runtime requirements, TP8 and TP4 environment profiles,
parser patches, and regression tests. Text-only serving has been validated on
eight NVIDIA RTX PRO 6000 Blackwell GPUs and on four NVIDIA B300 SXM6 GPUs. The
B300 lane successfully allocated and advertised a 524,288-token context with
the restored three-layer MTP draft. That result establishes deployment and API
compatibility; it is not a claim of semantic quality at a full 512K prompt.
The current kit also fixes two observed Step-3.5 tool-call failures: malformed
XML parameter closers leaking into arguments, and bare
{"name": ..., "arguments": ...} calls being returned as assistant text.
Both non-streaming and streaming parser paths are covered, and bare JSON is
converted only when its function name matches a tool offered in the request.
Intended access and limitations
If authorized, access is expected to be limited to approved users conducting controlled research, model evaluation, security testing, or deployment qualification. The candidate may produce inaccurate, unsafe, or otherwise unreliable output. Manual gating is an access-control workflow, not a safety guarantee. Independent safeguards, logging, human review, and compliance with applicable law remain the operator's responsibility.
License and disclaimer
Official StepFun source terms and derivative redistribution requirements remain under release review. The presence of gated model weights does not grant a license or authorize access, use, deployment, or redistribution. Any release will remain subject to verified source terms, required notices, and additional terms identified during authorization.
The information on this page is provided as-is and may change before release. For questions or responsible reporting, use this repository's Discussions.