Step 5 Preview Open Weights: What It Is and How to Use It Free

2026-10-16 · 4 min read

Step 5 Previewopen weightsfree accessmodel-release

StepFun opened the Step 5 Preview weights on October 15, 2026, and they are free to download under the StepFun Model License 1.0 — but free weights are not the same as free to run. The model has 600 billion total parameters, so serving it yourself takes a multi-GPU node; the hosted StepFun API is paid. aiwith has not integrated Step 5 Preview. Our closest options for long-context and coding work are Kimi K3 (1M-token context) and DeepSeek V4.1.

Last checked: October 16, 2026. This is an access guide, not a hands-on review.

Can I use Step 5 Preview for free?

Yes and no, and the difference matters:

The weights are free to download and reuse. The Hugging Face model card lists BF16 and FP8 checkpoints, also on ModelScope, plus an NVFP4 checkpoint built with NVIDIA.
Running it is not free unless you already have the hardware. StepFun's own deployment table starts FP8 serving at one node of 4 B300 or 8 H200 GPUs. With a GPU cloud bill attached, self-hosting is often more expensive than paying per token.
The StepFun hosted API is paid. There is no general "free Step 5 Preview chat" offer confirmed. The platform deducts from any free credit first, then your paid balance, so any temporary credit is worth checking on your own account.

If your goal is simply to use a strong model today, the weights being open does not lower the hardware bar. That is why the practical routes below are split between "run it yourself" and "pay per token."

What exactly was released?

StepFun describes Step 5 Preview as its flagship model for agentic work. According to the model card:

Architecture: sparse Mixture-of-Experts (MoE).
Size: 600B total parameters, 27B active parameters.
Context: up to 1 million tokens.
Input: text, images and video (per the developer docs).

One detail to plan around: the model card lists up to 1M output tokens, while the developer docs specify a 1M-token input window and a 64k-token maximum output. Use the smaller output figure when sizing a task.

StepFun markets particular strength in software engineering and in financial knowledge work. Those are vendor-reported claims; we have not benchmarked the model ourselves.

How do I use Step 5 Preview?

Pick the route that matches your hardware and budget:

Self-host with vLLM or SGLang. The model card publishes OpenAI-compatible serving commands for FP8 and BF16, including tensor, data and expert parallelism. You need a supported framework build with Step 5 Preview support.
Use the StepFun API. Request model id step-5-preview through their Open Platform with an API key.
Deploy the NVFP4 build on NVIDIA. The NVFP4 checkpoint targets Blackwell-class GPUs; StepFun and NVIDIA report roughly 8K tokens/sec/GPU on a GB300 in that configuration. Treat that as a vendor number on tuned hardware, not a general expectation.

All three routes are documented; none of them is a one-click install.

What does it cost?

For the hosted API, StepFun's pricing page lists Step 5 Preview per 1M tokens:

Input (cache miss): $1.00
Input (cache hit): $0.10
Output: $2.70

Output tokens include the model's reasoning as well as the final answer, so a reasoning-heavy task bills more than the visible reply suggests. For self-hosting, budget for the GPU nodes in the deployment table rather than a per-token rate.

What does "open weights" mean for the license?

The weights ship under the StepFun Model License 1.0, which the model card describes as Apache-2.0-style. In plain terms:

You may use, copy, distribute, modify and sublicense the model free of charge.
If a commercial product using it exceeds 100 million monthly active users or $20M in monthly revenue, you must display "Powered by StepFun AI" attribution.
Two cases require a separate commercial agreement: shipping the model's capabilities as a feature of hardware devices you manufacture or brand, and running a model-as-a-service business above $20M combined 12-month revenue.

For an individual developer or a small team, none of those thresholds apply today. If you plan to resell inference at scale, read the license before you build.

Can I use Step 5 Preview on aiwith?

No. As of this check, aiwith's model selector does not include Step 5 Preview. StepFun models are absent from our line-up, so an aiwith subscription does not unlock Step 5.

If you want a comparable open-weight or long-context model today, these aiwith options map to the same jobs:

Kimi K3 — also a 1M-token context model, useful for the long-document and agent tasks Step 5 targets.
DeepSeek V4.1 — the value option for coding and everyday work.
GPT-6 and Claude — separate frontier alternatives when you need a top proprietary model instead of open weights.

Open the model selector, confirm the version you selected, and check your allowance before starting. None of these is a renamed Step 5 Preview.