StepFun Step 5 Preview: 600B open-weight agent model

StepFun's Step 5 Preview, a 600-billion-parameter model built for AI agents, now has its weights on Hugging Face. StepFun promised the files for October 15, 2026, after a month in which the model was available only through its API. Developers can now download a large coding-focused model and run it themselves, though it needs a server full of data-center GPUs.

Weights are the trained numbers that make up a model. "Open weights" means anyone can download them, run the model locally and fine-tune it, within the license terms.

What Step 5 Preview is

The Hugging Face model card calls it "StepFun's flagship model for agentic work." It is a sparse mixture-of-experts (MoE) model. An MoE model splits its parameters into many smaller expert networks and uses only a few of them for each word, so it runs faster than its size suggests.

Spec	Step 5 Preview
Total parameters	600B (604B in the file metadata)
Active per token	About 27B
Context window	Up to 1M tokens
Input	Text and images
Weight formats	BF16 and FP8
License	StepFun Model License 1.0
Downloads	998 in the last month

The card describes the license as "an Apache-2.0-Style license" with extra attribution and commercial-license conditions. That differs from earlier reports. In October, Developers Digest noted that StepFun's announcement named no license and that the benchmarking firm Artificial Analysis listed the model as proprietary.

From placeholder to download

StepFun announced the model on September 20, with API access the same day. On that date its Hugging Face repository for the BF16 weights held a single file and no weights, OrcaRouter reported. "The repo is a placeholder, and that is the story," wrote Gideon Frost, the author of that post.

The model reached OpenRouter, a service that resells access to many AI models, on October 8. A day later, Developers Digest wrote that the weights were "promised for October 15, 2026." StepFun's GitHub repository for the model was updated on October 15. The model card now carries weight-format metadata and a download count.

Benchmarks and price

StepFun's own figures put it close to, but behind, the best closed models on coding tasks. The chart shows its DeepSWE v1.1 score next to rivals run at their "Max" settings, from the table Developers Digest reproduced. StepFun's announcement concedes that "a meaningful gap to the frontier remains."

The model card adds a 33.3 score on Terminal-Bench 4, a test of command-line tasks. On the independent Artificial Analysis Intelligence Index it scores 44. It ranked 24th of 200 models on September 18, per OrcaRouter, and 40th of 226 by October 9, per Developers Digest.

On OpenRouter it costs $1.00 per million input tokens and $2.70 per million output tokens. Developers Digest lists cached input at $0.05. Artificial Analysis also found it "very verbose": it produced 160 million output tokens on the index, against a median of 81 million.

What it takes to run

The model card lists example setups. In FP8, a compressed 8-bit format, vLLM runs it on four B300 GPUs or eight H200 GPUs in one server. In full BF16 it needs eight B300s, or 16 H200s across two servers. vLLM and SGLang are popular open-source programs for serving models.

That puts self-hosting out of reach for laptops and single workstations. It lands among other very large open releases this week, including Meta's 816B Muse Spark Open.

What this means for developers

• Read the license before you ship. The card's StepFun Model License 1.0 adds attribution and commercial-license conditions to an Apache-style base. Check whether your product triggers them before building on it.
• Budget for output, not just price. At $2.70 per million output tokens the rate is low, but the model writes a lot. Measure real tokens per task on your own workload, not the list price.
• Try the API before buying GPUs. Eight H200s is the smallest server setup listed. Test it through OpenRouter first, then decide whether self-hosting saves money.
• Treat vendor scores as a starting point. DeepSWE and StepCodeBench numbers come from StepFun. Run it on your own repositories and agent tasks before you switch.
• Use the recommended settings. The card suggests temperature 1.0, top_p 0.95 and top_k 0, with results reported at the "High" reasoning setting. Different settings may not match the published scores.