LLM Releases
← Catalog

Step 5 Preview

Preview
StepFunFrontierProprietary

Step 5 Preview is StepFun's flagship agentic model, launched in preview via the StepFun API and Studio on 2026-09-20 (Artificial Analysis dates it 2026-09-18). A 600B-total / 27B-active sparse Mixture-of-Experts in a narrow-deep 92-layer stack (~4.5% of parameters active per token), it accepts text plus image (up to 60 per request) and video input and emits text, over a 1,000,000-token context with up to 64,000 output tokens. Positioned for long-horizon agentic work across large codebases and documents (tool use with multi-step refinement) with particular strength in software engineering and finance. Served through the StepFun API at $1.00 / 1M input and $2.70 / 1M output tokens (cache reads $0.05 / 1M; reasoning tokens billed as output). Open weights are promised for 2026-10-15 (the Hugging Face repo stepfun-ai/Step-5-Preview-BF16 held no weights or license as of 2026-10-08), so this is API-only for now. StepFun-reported benchmarks include GPQA Diamond 93.5%, Humanity's Last Exam 46.5%, DeepSWE v1.1 67.7%, SWE-Marathon v1.1 72.7%, and StepCodeBench (StepFun's own) 49.0%; Artificial Analysis independently measured Intelligence Index v4.3.2 = 44 (24th of ~200 models) and Terminal-Bench 4.0 = 33.3%.

Specifications

License
Proprietary · Proprietary (open weights promised 2026-10-15)
Weights
Not released
Architecture
Mixture-of-Experts
Parameters
600B · 27B active
Context window
1M tokens
Max output
64K tokens
Knowledge cutoff
—
Price (in / out, $/M)
$1 / $2.7
Modalities
TextVisionVideoCode

Benchmarks

No benchmark scores recorded yet. Spotted some? Submit a correction.

Vendor-reported figures are claims until independently verified. See methodology.