Step 5 Preview is an open-weight sparse mixture-of-experts model. The Hugging Face card says 600B total and 27B active parameters, with context up to 1M tokens. StepFun’s deployment examples are multi-GPU. The card says actual memory, usable context, and concurrency vary, and it does not give one VRAM number. This site does not invent one and does not add discrete GPUs into a single pool.
The public product page at stepfun.com did not include a spec table when checked. Dates below are repo timestamps and document checks, not a claimed open-weights holiday. Checked 16 October 2026.
Specs that disagree, both quoted
| Item | Model card | Platform docs |
|---|---|---|
| Model | Step 5 Preview | API id step-5-preview |
| Context | Up to 1M tokens | 1M tokens. The vLLM notes say the 1M limit covers input and output combined |
| Max output | Up to 1M tokens | 64k tokens |
| Input | Intro: text, images, and video. Table: text, image | Text, images, and video |
| Output | Text | Text |
| License | StepFun Model License 1.0, described on the card as Apache-2.0-style with extra attribution and commercial conditions | Not restated |
Card: stepfun-ai/Step-5-Preview. The license file is LICENSE.txt. This page does not paraphrase clauses that were not read beyond the card’s own summary. Docs: model page, last modified 10 October 2026 in that page’s metadata.
Safetensors metadata total on the FP8 and BF16 repos is 604,339,504,832. The card’s round number is 600B. The NVFP4 repo’s metadata total is 320,612,382,016, which is a packed count, not 600B.
Deployment examples
From the card. B300, H200, and the other accelerators below are not pages in this catalog.
| Framework | Precision | Hardware on the card |
|---|---|---|
| vLLM | FP8 | One B300 node, 4 GPUs |
| vLLM | FP8 | One H200 node, 8 GPUs |
| vLLM | BF16 | One B300 node, 8 GPUs |
| vLLM | BF16 | Two H200 nodes, 16 GPUs |
| Transformers | FP8 | One H200 node, 8 GPUs |
| Transformers | BF16 | Two H200 nodes, 16 GPUs |
| SGLang | FP8 | One B300 node, 4 GPUs, or one H200 node, 8 GPUs |
The card says BF16 examples do not guarantee memory for a 1M-token request. Setting a BF16 dtype on the FP8 checkpoint does not convert the weights. See two GPUs and unified memory.
File bytes, by component
GB is 10⁹ bytes. Sums are download size, not resident VRAM. See file size is not runtime VRAM.
| Component | Repo and revision | Bytes | GB |
|---|---|---|---|
model-00001 through model-00024 |
FP8 364b3b03fed2 |
635,107,017,592 | 635.107 |
model-00001 through model-00024 |
BF16 7734b7ffac11 |
1,199,690,207,296 | 1,199.690 |
model-mtp3-full-00001.safetensors |
FP8 and BF16, same size | 5,031,847,968 | 5.032 |
model-mtp3-00001.safetensors |
FP8 and BF16, same size | 2,246,149,176 | 2.246 |
vit.safetensors |
FP8 and BF16 | 3,962,113,752 | 3.962 |
model-vit-00001.safetensors |
FP8 and BF16 | 3,962,113,720 | 3.962 |
The two vision files differ by 32 bytes. They are listed separately and are not added. The two MTP files have different sizes. The card does not say that a serving command loads both, so they are not folded into the main shard total. NVFP4 revision bc246cc2848e uses a different layer layout. Its safetensors files other than model-vit-000.safetensors sum to 390,203,088,560 bytes (390.203 GB). That vision file is 3,962,113,752 bytes. Repo createdAt for the FP8 and BF16 repos is 20 September 2026. That timestamp is not labeled as the open-weights announcement.
API price
Pricing details, per 1M tokens for step-5-preview:
| Cache-miss input | Cache-hit input | Output |
|---|---|---|
| $1.00 | $0.10 | $2.70 |
The cache-miss price includes writing new content to the cache. Output tokens include reasoning and the final answer.
What this page does not claim
No calculator preset. Qwen3.8-27B is an open model this site does estimate. Nemotron 3.5 Super VL is another open-weight model whose publisher lists multi-GPU targets instead of one VRAM number. The formula scope is the methodology.
FAQ
How much VRAM does Step 5 Preview need?
StepFun does not publish a single VRAM number. The model card says memory, usable context, and concurrency vary with the checkpoint, framework, and KV cache. Its examples include vLLM FP8 on 4 B300 GPUs or 8 H200 GPUs, and vLLM BF16 on 8 B300 GPUs or 16 H200 GPUs across two nodes. Those GPUs are not added into one memory pool on this site.
How large are the Step 5 Preview weight files?
On 16 Oct 2026 the FP8 repo's model-00001 through model-00024 shards sum to 635,107,017,592 bytes (635.107 GB). The BF16 repo's matching shards sum to 1,199,690,207,296 bytes (1,199.690 GB). Vision and MTP files are extra and are not included. The NVFP4 safetensors metadata total is 320,612,382,016, which is not the card's 600B parameter count.
What is the Step 5 Preview context length and price?
The card says context up to 1M tokens and maximum output up to 1M tokens. The platform docs say the context window is 1M, maximum input is 1M, and maximum output is 64k tokens. API id step-5-preview is priced per 1M tokens at $1.00 input cache miss, $0.10 input cache hit, and $2.70 output. Output tokens include reasoning and the final answer.
Does Step 5 Preview accept video?
The card introduction says text, images, and video. The card specification table says input is text and image, and output is text. The platform docs say input types are text, images, and video, and output is text. Both documents were checked on 16 Oct 2026. This page does not pick a winner.