Models & LLMs

Step 5 Preview: Open Weights, Price and Benchmarks

StepFun's Step 5 Preview weights are now on Hugging Face under the StepFun Model License 1.0. Here is what the license, API prices, limits and independent scores mean for piloting or self-hosting.

Toolbit AI - Team
5 min read
Step 5 Preview: Open Weights, Price and Benchmarks

StepFun released the Step 5 Preview open weights on Hugging Face on October 15, in BF16 and NVFP4, under the StepFun Model License 1.0. StepFun's docs price the API at $1.00 per 1M input tokens, $0.10 for cache hits and $2.70 for output. Artificial Analysis, the independent benchmarking service, scores it 44 on its Intelligence Index.

At the September launch, the API was the only way in and no license was named. Now you can choose between calling the API and running the weights yourself, if the license terms fit your product.

  • What changed: weights are live and the license has a name.
  • Who's affected: teams piloting the API, and teams that waited for weights to self-host.
  • What to do: read the license, then use the decision rule at the end of this post.

What changed since the September launch?

Two things have changed since the September launch: the weights are downloadable, and the license is named.

At launch (September)Now (checked October 21, 2026)
WeightsPromised for October 15Live on Hugging Face in BF16 and NVFP4
LicenseNone namedStepFun Model License 1.0

Artificial Analysis now labels Step 5 Preview an "Open weights model" and lists the StepFun Model License 1.0 and Hugging Face as its weights source. The score is still 44. AA now ranks it #4 of 121 among open-weight large models; the #26 of 202 overall rank from launch week is no longer shown, so the two ranks aren't comparable.

What can you download, and under what license?

You can download the full model now from the stepfun-ai organization on Hugging Face, which lists three repos: Step-5-Preview, Step-5-Preview-BF16 and Step-5-Preview-NVFP4. The model card describes a sparse mixture-of-experts model with 600B total and 27B active parameters, 1M tokens of context, and text, image and video input.

The license is the StepFun Model License 1.0. The model card's description: "an Apache-2.0-Style license, with additional attribution and commercial-license conditions specific to model use, modification and distribution." Hugging Face's own license field just says "other".

Read LICENSE.txt before you ship anything. Third-party write-ups quote a usage threshold that triggers an attribution notice. That figure is not confirmed against the license text, so treat it as unverified.

Plan the hardware before you plan the pilot. In BF16, 600B parameters at 2 bytes each is about 1.2 TB of weights before any KV cache. Hugging Face labels the NVFP4 repo 321B against 604B for BF16, so it is the lighter download if your inference stack supports the format. StepFun recommends temperature 1.0, top_p 0.95 and top_k 0.

What does the Step 5 Preview API cost?

StepFun's pricing page lists step-5-preview at $1.00 per 1M input tokens on a cache miss. Cache hits cost $0.10 and output costs $2.70 per 1M tokens. The cache-miss price already includes writing new content to the cache.

For scale, StepFun's older and cheaper step-3.7-flash is listed on the same page:

Price per 1M tokensstep-5-previewstep-3.7-flash
Input, cache miss$1.00$0.20
Input, cache hit$0.10$0.04
Output$2.70$1.15

The cache is where agent loops save money, at a 90% discount per re-read. An agent that reads a 1M-token repository 10 times pays $1.00 for the first read and $0.10 for each of nine re-reads.

Reading a 1M-token repo 10 times costs $10.00 in input without cache hits and $1.90 with them, 81% less

That is $1.90 with cache hits against $10.00 without, 81% less. A single call with 1M input and 0.1M output tokens and no cache hit costs $1.27.

Two things push the real bill up. Artificial Analysis rates the input price "expensive" against a $0.43 median and the output price "somewhat expensive" against a $1.68 median. It also counted 160M output tokens across its Intelligence Index run, about 14% above the 140M median. Its cost per index task is $1.03. Verbose answers cost more when every output token is $2.70 per million.

Which benchmark numbers are independent?

Only the Artificial Analysis figures are independent. It scores Step 5 Preview 44 on its Intelligence Index, #4 of 121 in its comparison set. Output speed is 91.8 tokens per second, #28 of 78.

StepFun's coding and reasoning scores come from its own runs, reported on the model card with the High reasoning setting. Those figures could not be re-verified for this article, so they are left out here. Treat any launch-week benchmark you see quoted as a vendor claim until someone reproduces it on open weights. For an open-weight rival on the same kind of footing, see our coverage of GLM-5.3.

What limits and gaps should you plan around?

The API caps output at 64k tokens, while input goes to 1M. The Hugging Face model card lists "Up to 1M tokens" of maximum output for the weights. The two sources differ, so plan for 64k on the API and test longer outputs only on your own deployment.

Media input on the API allows up to 60 images per request in JPG, PNG, WebP or static GIF. Video can be MP4, QuickTime or Matroska, with each MP4 under 128 MB and under five minutes recommended. Artificial Analysis lists only text and image input, so its score does not cover video.

Rate limits depend on your top-up tier. StepFun's pricing page lists these for individual accounts:

TierTop-upConcurrencyRequests per minuteTokens per minute
V0Under $155100500,000
V4$1,500 and above1302,60013,000,000

Five concurrent requests at V0 is enough for a pilot, not for a fleet of parallel agents.

Self-host or API: which should you pick?

Pick by output length, hardware and license fit, in that order.

  • Call the API if you want results this week, your outputs fit in 64k tokens and your workload re-reads the same context, where cache hits cost $0.10.
  • Self-host if you can run a 600B mixture-of-experts model (about 1.2 TB in BF16), need data on your own infrastructure or longer outputs, and your legal team accepts the license's attribution and commercial conditions.
  • Look elsewhere if price per token decides it. StepFun's own step-3.7-flash costs $0.20 in and $1.15 out, and the wider slide in inference costs gives you other options.

Next step: replay tasks you already run, for example from a Claude Code workflow, against step-5-preview at the V0 tier. Log output tokens per task, since verbosity drives the bill, then decide between staying on the API and moving to the weights.

Share this article

Related articles

Continue exploring similar guides and insights