Step 5
A large language model developed by StepFun, supporting multimodal understanding, code generation, and logical reasoning. The model offers multiple reasoning modes (low, medium, and high) and is designed to be more compact than comparable flagship models. It follows an open-source-friendly tradition under the Apache 2.0 license.
Timeline (last 90 days)
A preview version of StepFun's Step 5 appeared on the model platform OpenRouter during the same week.
The developer used a free trial week of a Step 5 model on OpenRouter to test large-scale context dumping, and the model was able to correctly parse a context window of 1 million tokens while maintaining tool-calling integrity.
Step 5's model size is significantly smaller than comparable flagship competing products.
The Step 5 model was only available for approximately one hour on its release day before being taken down, with an official launch planned for next month.
Step 5 frequently triggers runtime errors on some complex tasks under Q4 quantization, requiring a higher quantization precision such as Q5 or above for stable output.
Step 5 officially provides three levels of thinking modes — low, medium, and high — and is not a brute-force reasoning model.
StepFun AI (阶跃星辰) has launched the Step 5 preview model.
Step 5 performed consistently in logic and trick question tests, correctly answering the orange distribution problem and the surgeon trick question without falling into training bias.
Step 5 will refuse to generate code related to nuclear weapons, with the input being deemed invalid.
Step 5 performed poorly on tasks such as GTA-style city generation, kart racing, and 3D human anatomy, with issues including roads passing through walls and incorrect heart positioning.
3 more timeline events
All Facts (10)
Step 5 performed poorly on tasks such as GTA-style city generation, kart racing, and 3D human anatomy, with issues including roads passing through walls and incorrect heart positioning.
50%UnverifiedStep 5's model size is significantly smaller than comparable flagship competing products.
50%UnverifiedStep 5 frequently triggers runtime errors on some complex tasks under Q4 quantization, requiring a higher quantization precision such as Q5 or above for stable output.
50%UnverifiedStep 5 officially provides three levels of thinking modes — low, medium, and high — and is not a brute-force reasoning model.
50%UnverifiedStep 5 has multimodal capabilities, being able to identify an orange tabby cat in an image and generate Three.js 3D models and voxel-style models.
50%UnverifiedStep 5 performed consistently in logic and trick question tests, correctly answering the orange distribution problem and the surgeon trick question without falling into training bias.
50%UnverifiedStep 5 will refuse to generate code related to nuclear weapons, with the input being deemed invalid.
50%UnverifiedA preview version of StepFun's Step 5 appeared on the model platform OpenRouter during the same week.
50%UnverifiedThe developer used a free trial week of a Step 5 model on OpenRouter to test large-scale context dumping, and the model was able to correctly parse a context window of 1 million tokens while maintaining tool-calling integrity.
50%UnverifiedStepFun AI (阶跃星辰) has launched the Step 5 preview model.
50%