Step 5 Preview: Can You Actually Run This 600B Model?

Learn more：https://treerouter.com

Abstract

On October 15, 2026, StepFun opened the weights of Step 5 Preview, its flagship agent-oriented model. The total parameter count is roughly 600 billion, with 27 billion active parameters per token and a native context window of one million tokens. Weights are distributed through Hugging Face and the ModelScope community in two checkpoints: an FP8 variant of about 650 GB and a BF16 variant of about 1.2 TB. The license is StepFun Model License 1.0, a proprietary agreement with attribution and commercial-authorization conditions rather than standard Apache-2.0. Official benchmark claims position it ahead of Kimi K3 on all eight reported metrics, though each individual score is surpassed by at least one closed-source flagship. This article examines the open-sourced artifacts, the sparse-MoE architecture, training disclosures, the exact license obligations, the eight-item benchmark table, and the deployment matrix for vLLM, SGLang and Transformers.

1. What Was Open-Sourced

Step 5 Preview is positioned in the model card as StepFun's flagship model for agent work, focusing on repository-level software engineering, frontend and visual development, device programming, long-horizon agent execution, professional knowledge work and finance.

The official sizing is "600B total parameters, 27B active". Hugging Face safe-tensor statistics sum to 604.3 billion, meaning the 600B figure is a rounding. Within the FP8 repository, roughly 564.7 billion parameters are stored in FP8 while about 39.6 billion remain in BF16.

Two checkpoints appeared simultaneously on October 15, 2026:

• FP8: stepfun-ai/Step-5-Preview, about 650 GB of total storage, split across 28 safetensors shards.
• BF16: stepfun-ai/Step-5-Preview-BF16, about 1.2 TB of total storage.

Neither repository sets up an access-application gate on Hugging Face. The GitHub repository stepfun-ai/Step-5-Preview currently contains only the README, Chinese and English license files, a contributor list and image assets. It does not include model weights.

2. Architecture: A Narrow-and-Deep Sparse MoE

Step 5 Preview is a 92-layer sparse Mixture-of-Experts model. Each token is routed through 8 of 352 routed experts, plus one shared expert.

Layer structure: of the 92 layers, 88 are MoE layers and 4 are dense layers. The hidden dimension is 4096, with 64 attention heads and 4 KV groups.

Attention layout follows a 3:1 alternation inherited from Step 3.5. In every group of 4 layers, 3 use sliding-window attention with a window of 512 tokens, and 1 uses full attention.

Sparse GQA: within full-attention layers, 8 tokens are merged into one block, and top-k selects 512 blocks, equivalent to 4096 tokens. The blog states that with a block size of 8, the indexer and top-k overhead drop to one-eighth of the original, with a token coverage of 78.06 percent, compared with 79.09 percent for the per-token baseline configuration at k=2048.

Multi-token prediction: the configuration includes 3 layers of MTP, corresponding to the vLLM startup flags --spec-method mtp --spec-tokens 3.

Vision: the model uses a 47-layer visual encoder with 728-pixel input and a patch size of 14.

StepFun summarizes the philosophy as "narrow and deep". The 92-layer depth is comparable to very large open models such as Qwen 3.8 Max and Kimi K3, yet the total parameters and active parameters are respectively less than one-quarter and one-third of those models. This is StepFun's own framing and should be read as a self-reported claim.

3. Training Details Disclosed in the Technical Blog

The technical blog describes training stability and agent reinforcement-learning engineering in concrete terms, but it provides no benchmark numbers and does not state the open-source scope or license terms.

Several engineering choices stand out:

• FP32 hidden states: hidden states and their gradients are kept in FP32, reducing peak memory by nearly an order of magnitude with small extra overhead. The blog acknowledges the underlying reason "remains unclear".
• FP32 normalized scaling parameters: with Zero-centered RMSNorm, when gamma approaches -1, BF16 precision becomes insufficient and channels oscillate on and off.
• Fine-grained gradient clipping: only about 0.00004 percent of gradient values are clipped per layer.
• Training-inference alignment: at a staleness of zero, training and inference align bit-for-bit; TP=1 training matches TP=8 inference with overhead below 10 percent.
• Data and environment: four components synthesize the data, described as "millions of difficult, verifiable environments". Context compression and task rewards are trained end-to-end.
• RL system efficiency: the blog reports MTP-3 delivers about 30 percent end-to-end speedup, native FP8 MoE is about 20 percent faster, and effective KV-cache capacity expands by 8 to 12 times.

4. License: Not Apache-2.0, and Where Attribution or Separate Authorization Applies

Step 5 Preview uses StepFun Model License 1.0. The model card calls it an "Apache-2.0-style license with additional attribution and commercial-authorization conditions". The Hugging Face license tag reads other, and the GitHub license field shows NOASSERTION.

The base grant is perpetual, worldwide, non-exclusive, free and royalty-free. It permits use, copying, distribution, sublicensing, modification and creation of derivative works, and includes a patent grant. Several restrictions are easy to overlook.

Derivative-work scope is broad. Fine-tuning, distillation, merging, quantization, pruning and continued pretraining all produce derivative models. Models trained substantially on the model's outputs also count as derivatives. A standalone application that only calls the model via API is not treated as a derivative work on that basis.

Redistribution requires attaching the license, retaining notices, marking modifications, and providing the model card or a link to it.

Attribution applies when, in any calendar month, a licensee or its affiliate's commercial product exceeds 100 million active users, or revenue exceeds 20 million US dollars. In that case, the interface (including the About page) and the terms of service or product documentation must display "Powered by StepFun AI".

Separate commercial authorization is required in two scenarios. First, hardware manufacturers or brands shipping phones, cars, smart glasses, wearables, robots or smart-home devices must obtain authorization before first shipping the capability as a device feature. Second, a licensee operating an MaaS business with combined revenue exceeding 20 million US dollars over any consecutive 12 months requires authorization for any commercial use.

The agreement provides two exceptions for the separate-authorization clause: internal use that does not provide the model, derivatives, outputs or capabilities to third parties; and use through StepFun-certified channels within the certified scope.

MaaS is defined as providing inference or fine-tuning access (including APIs) where the third party holds substantial control over inputs, parameters or training data. Embedding a capability into a specific end-user function, or merely forwarding requests, does not count as MaaS.

Licensees are responsible for compliance. The agreement grants no trademark rights. Serious breach terminates the license, recoverable by correcting within 30 days of written notice. The agreement exists in Chinese and English versions, with the English version controlling. Questions can be directed to the listed official email license@stepfun.ai.

These points are a restatement of the terms, not legal advice.

5. Official Benchmark Table: Comparison with Kimi K3 and GLM-5.3

The model card's chart compares Step 5 Preview with Kimi K3, GLM-5.3 and two closed-source flagships. Across all eight metrics, Step 5 Preview leads Kimi K3, but each metric trails at least one closed-source flagship.

Three takeaways from the table:

• Against Kimi K3: all eight metrics lead.
• Against GLM-5.3: six lead, while Terminal-Bench v4 (33.3 vs 41.9) and GDPval-AA v2.1 (1566 vs 1645) trail.
• Against the two open flagships: DeepSWE, StepCodeBench, ProgramBench and Terminal-Bench v4 trail both; the remaining four beat only one, and none lead both simultaneously.

Caveats before citing these numbers: the source is the official model card chart, which is vendor-reported, and this text transcribes the figures with the original chart as authority. The Step 5 Preview row uses the High reasoning setting, while the other models are labeled Max, so the two sides are not configured identically. GDPval-AA data comes from Artificial Analysis as of September 20, 2026; DeepSWE v1.1 uses the SWE-agent harness with temperature 1.0 and top_p 0.95.

6. Deployment: Self-Host or Call the Official API

Two paths are available now. FP8 weights are about 650 GB, and the self-host starting threshold is a full-node GPU allocation. Calling the official API has a lower entry barrier for quick evaluation.

Option One: Deploy the Weights

The model card lists the following starting configurations.

The model card also warns that actual memory usage, available context and concurrency depend on the checkpoint, framework and KV-cache allocation. The BF16 configuration is not guaranteed to fit a 1-million-token request. Adding --dtype bfloat16 to an FP8 checkpoint will not convert the weights to BF16; a separate BF16 checkpoint is required.

Step 1, pull the official vLLM image and start the service with the 4-card B300 TP4 form:

docker pull vllm/vllm-openai:step-5-preview
docker run --rm -it --gpus all --ipc=host --network=host \
  --volume "$HOME/.cache/huggingface:/root/.cache/huggingface" \
  --entrypoint /bin/bash vllm/vllm-openai:step-5-preview
# inside the shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
vllm serve stepfun-ai/Step-5-Preview \
  --host 0.0.0.0 --port 18000 --served-model-name "step5" \
  --tensor-parallel-size 4

For an 8-card H200, change CUDA_VISIBLE_DEVICES to 0 through 7 and set --tensor-parallel-size 8. To enable expert parallelism, add --enable-expert-parallel; on B300 also add --moe-backend deep_gemm_mega_moe, while H200 does not need this parameter.

Step 2, send a request:

curl [http://127.0.0.1:18000/v1/chat/completions](http://127.0.0.1:18000/v1/chat/completions) \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "step5",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 4096,
    "stream": true
  }'

The recommended sampling parameters from the model card are temperature=1.0, top_p=0.95 and top_k=0.

Step 3 covers insufficient memory or switching frameworks. For memory pressure, first reduce --max-model-len or concurrency. To switch to SGLang, use the lmsysorg/sglang:step-5-preview image; the full startup arguments are in the model card and the linked SGLang Cookbook. To switch to Transformers, use the scripts under the inference/ directory, run convert.py --check-only to inspect the checkpoint, then convert and call generate.py with torchrun. The officially validated combination is PyTorch 2.13.0 with Transformers 5.15.0. The script loads the local tokenizer.json directly, because Transformers 5.15's automatic tokenizer-class selection is incompatible and can break long prompts.

Option Two: Call the Official API

The official API is OpenAI-compatible. The base address is [https://api.stepfun.ai/v1](https://api.stepfun.ai/v1), with the conversation endpoint /chat/completions. The model ID is step-5-preview. Keys are created on the Interface Keys page of the open platform and passed via an Authorization: Bearer header.

Reasoning intensity uses reasoning_effort in Chat Completions and output_config.effort in the Messages API, with values low, medium and high. Maximum output is 64k tokens. Streaming, tool calling, JSON Mode, JSON Schema and prompt caching are supported.

curl [https://api.stepfun.ai/v1/chat/completions](https://api.stepfun.ai/v1/chat/completions) \
  -H "Authorization: Bearer YOUR_STEP_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "step-5-preview",
    "reasoning_effort": "low",
    "messages": [{"role": "user", "content": "Explain sparse MoE in three sentences."}]
  }'

7. Pricing, Rate Limits and Caveats

The official pricing page lists step-5-preview at 1 USD per million input tokens on a cache miss (already including write cache), 0.1 USD per million tokens on a cache hit, and 2.7 USD per million output tokens. Pricing is not tiered by context length. Output tokens include both the reasoning process and the final answer. At these rates, a request with 100,000 input tokens and 5,000 output tokens with no cache hit costs roughly 0.11 USD.

Account rate limits are tiered by cumulative recharge. V0 (cumulative recharge below 15 USD) allows concurrency 5 and 100 requests per minute; the highest tier V4 allows concurrency 130 and 2,600 requests per minute.

Several documentation inconsistencies require source attribution:

• Maximum output: the model card spec table writes "maximum output up to 1M", while the official API documentation writes 64k. The model card's deployment notes treat 1M as the combined input-plus-output upper bound.
• Input types: the introduction and official documentation mention video, while the model card spec table lists only text and image.
• License: the model card says "Apache-2.0-style", but the actual agreement is a proprietary license with additional conditions and cannot be read as standard Apache-2.0.
• Technical blog: it contains no benchmark numbers and does not state open-source scope or license terms; these must be taken from the model card and license file.
• Maturity: the project remains marked Preview, and the blog itself notes the related systems are "still evolving".

8. FAQ

Can Step 5 Preview be used commercially?
The license does not generally prohibit commercial use. Three conditions apply: products exceeding 100 million active users or 20 million USD revenue must attribute "Powered by StepFun AI"; hardware manufacturers shipping it as a device feature need separate authorization; and MaaS operators with combined revenue over 20 million USD over any consecutive 12 months require separate commercial authorization.

How many GPUs are needed for local deployment?
FP8 with vLLM requires one 4-card B300 node or one 8-card H200 node. BF16 requires one 8-card B300 node or two 8-card H200 nodes, and even then is not guaranteed to fit a 1-million-token request.

What is the API price?
Input cache-miss is 1 USD per million tokens, cache-hit is 0.1 USD per million tokens, and output is 2.7 USD per million tokens, per the official pricing page.

Can reasoning be disabled?
The documented reasoning intensities are only low, medium and high. No off option exists. Output tokens include the reasoning process and are billed accordingly.

Can it process images and video?
Per official documentation, input can be text, image and video. Up to 60 images per request (JPG/JPEG, PNG, WebP, static GIF). Video supports MP4, QuickTime and Matroska; a single MP4 passed by URL must be under 128 MB, with a suggested duration under 5 minutes. Output is text only.

9. Conclusion

Step 5 Preview's open-weights release gives engineering teams a flagship-grade sparse-MoE agent model they can run on their own hardware or consume via API. The narrow-and-deep 92-layer design, Sparse GQA, three-layer MTP and vision encoder together target long-horizon agent and coding workloads. The proprietary license, however, carries real obligations: broad derivative-work coverage, an attribution threshold and two separate-commercial-authorization scenarios that hardware vendors and large MaaS operators must review carefully before deployment.

The deployment baseline is high. FP8 weights of about 650 GB demand a full B300 or H200 node, and BF16 roughly doubles that requirement while still not guaranteeing a one-million-token request fits. Teams evaluating the model should confirm their runtime framework versions against the official matrix and treat the vendor-reported benchmark figures with appropriate caveats.

As of October 15, 2026, the information above derives from StepFun's model card, license file, technical blog, official documentation and pricing page. The project remains Preview; parameters, prices and terms may change.

References

• Hugging Face (FP8): https://huggingface.co/stepfun-ai/Step-5-Preview
• Hugging Face (BF16): https://huggingface.co/stepfun-ai/Step-5-Preview-BF16
• ModelScope: https://modelscope.cn/models/stepfun-ai/Step-5-Preview