StepFun Step 5 Preview open weights: 600B parameters, a conditional license and a verbosity problem

StepFun's Step 5 Preview is a sparse mixture-of-experts model with 600B total parameters and 27B active per token, and its weights are now on Hugging Face. The model card lists a 1M-token context and a StepFun Model License 1.0 with attribution and commercial conditions. StepFun's own coding scores are mid-pack, and an independent index run flagged the model as verbose.

What is on Hugging Face

The model card describes "a sparse Mixture-of-Experts (MoE) model with 600B total parameters and 27B active parameters." The safetensors metadata lists 604B parameters, and there is a smaller NVFP4 variant at 321B. The context window is up to 1M tokens, with output of up to 1M tokens too. The specification table lists text and image input and text output, though the introduction also mentions video, so the card is inconsistent on that point. The page does not give a release date. A Developers Digest write-up from October 9, 2026 reported that StepFun said the weights would ship on October 15, and the repository shows a recent update, so the release landed on or near that date.

If you are new to why total and active counts differ, our mixture-of-experts explainer covers the router and the trade-offs. In short, only 27B parameters do work on each token, which keeps compute per token far below a dense 600B model. All 600B still have to sit in memory.

The license is the first thing to read

The card names the StepFun Model License 1.0 and describes it as Apache-2.0-style with additional attribution and commercial-license conditions. That is not the same as Apache 2.0 or MIT. If you plan to build a product on the weights, the specific attribution duties and the commercial-license trigger decide whether this is usable. The summary on the card does not state the trigger, so a legal read of the full text is the next step. Until then, call this open weights, not open source.

Benchmarks: what StepFun reports

Evaluation results on the card, run on the High reasoning setting, include:

• DeepSWE (datacurve/deep-swe): 67.7
• Agents-last-exam, CLI pass rate: 29.5
• Terminal-bench (terminalbench_4): 33.3

The full benchmark table is an image, so most competitor comparisons are not in the page text. The Developers Digest summary adds that StepFun reports ProgramBench at 80.5 and a Terminal-Bench v4 score behind GLM-5.3's 41.9, and that StepFun compares its High setting against rivals' Max settings. That last point deserves attention. Comparing a mid reasoning setting to a rival's top setting can flatter or hurt depending on direction, and readers cannot tell from a headline number.

StepFun is also candid on the gap. The announcement, as quoted by Developers Digest, concedes that "a meaningful gap to the frontier remains" on the hardest long-running tasks. For context, DeepSeek's card for its newest Flash release lists Terminal-Bench 4.0 at 42.3 and DeepSWE v1.1 at 77.0, though harnesses and settings differ, so treat that as a rough comparison only. We cover that release separately.

The independent score and the verbosity flag

Artificial Analysis gave Step 5 Preview an Intelligence Index of 44 as of the October 9 snapshot, against a median of 26 for comparable reasoning models, according to Developers Digest. The same summary says the page calls the model very verbose: 160M output tokens for the index run, against a median of 81M. I could not open the Artificial Analysis page directly this run, so the figures are secondhand and may have moved since. Different trackers have quoted different ranks and medians for the same model, which is a reminder to check the live leaderboard. Our piece on what leaderboards rank explains why.

Verbosity matters because most API pricing is per output token. A model that writes twice the median length pays that back in cost and latency. Developers Digest quotes OpenRouter pricing of $1.00 per million input tokens and $2.70 per million output tokens on the StepFun provider as of that date, and early telemetry showed structured-output errors around 17%. Those are one provider's moving numbers. Test your JSON schemas before relying on it.

What it takes to run

The card's deployment table is useful. For vLLM at FP8, you need one B300 node with 4 GPUs or one H200 node with 8 GPUs. At BF16 it needs one B300 node with 8 GPUs, or two H200 nodes with 16 GPUs, and the page warns that these setups do not guarantee memory for a full 1M-token request. SGLang at FP8 has the same footprint as vLLM at FP8. Recommended settings are temperature 1.0, top_p 0.95 and top_k 0. This is a data center model, not a workstation one.

How to test it before you commit

A fair bake-off needs three controls. First, fix the reasoning setting, because StepFun's own table compares its High setting with rivals' Max settings and the gap could flip with a matched setting. Second, count total tokens per solved task, not tokens per second, since the verbosity flag means output length drives your bill. Third, run the same harness for every model, because Terminal-Bench and DeepSWE results shift with the agent scaffold around the model. If Step 5 Preview wins on your repository under those conditions, the license review is worth the legal time. If it only wins on its own announcement table, it has not yet earned a place in your stack.

Our take

Step 5 Preview is a credible addition to the open-weight field and a candid one, given StepFun's own admission of a frontier gap. It is not a clear winner. The license has conditions, the coding scores trail GLM-5.3 on Terminal-Bench, and the verbosity flag undercuts the low per-token price. If you already run large open models on B300 or H200 nodes, it is worth a bake-off on your own tasks, measured in cost per solved task. If you are choosing a first open model for coding agents, the permissive licenses and stronger published scores elsewhere make this a second-round candidate.