StepFun: Step 5 Preview

StepFun

balanced

Capabilities

VisionToolsStructured OutputExtended ThinkingLong Context

Strengths

Exceptional long-context handling (1M token window)
Efficient sparse MoE design reduces latency relative to dense flagships
Strong multimodal capability (text, image, video)
Open-weight model enables self-hosting and fine-tuning

Weaknesses

Very slow inference speed (0.06 score; ~88 tok/s) limits real-time agentic work
Limited public methodology and training data documentation
Sparse MoE complexity may reduce compatibility with some inference frameworks
Early preview status; stability and feature maturity uncertain

Pricing

Input / 1M tokens

Output / 1M tokens

$2.7

Context window

1,000k

Transparency

Open weights

10.0 / 10

Open training data

0.0 / 10

Open methodology

6.5 / 10

Licence openness

7.0 / 10

Provider disclosure

5.5 / 10

FMTI company score

N/A

Composite:5.5 / 10

Open-weight sparse MoE architecture with publicly available weights, but limited training methodology and data transparency. Pricing and capability disclosures are moderate.

Sustainability

Inference energy

N/A

Training footprint

N/A

Provider infrastructure

5.0 / 10

Composite:5.0 / 10

Sparse MoE (27B active parameters) reduces inference energy vs. dense models of similar scale, but provider sustainability commitments and infrastructure details are not well-documented.

Task Fitness

question answering

75%

code

80%

mathematics

50%

business communication

50%

vision

75%

analysis

75%

extraction

80%

long-form generation

75%

research

50%

reasoning

75%

summarisation

80%

translation

75%

conversation

75%