StepZen's Open-Source AI Model Takes Second Place Globally
StepZen's Breakthrough AI Model Goes Fully Open-Source
In a significant move for the AI community, StepZenith announced on March 4th the complete open-sourcing of its Step3.5Flash model. This includes not just the final product but also intermediate training weights and the proprietary Steptron training framework - giving developers unprecedented access to cutting-edge AI technology.
Built for Efficiency and Power
The Step3.5Flash model employs an innovative sparse MoE (Mixture of Experts) architecture totaling 196 billion parameters. What makes this particularly impressive is its energy efficiency - during actual operation, it activates only about 11 billion parameters at a time. This smart design allows it to achieve blazing speeds of up to 350 transactions per second on coding tasks while maintaining excellent performance on complex reasoning challenges.
Rapid Adoption in Developer Community
The tech world has responded enthusiastically to this release. On Hugging Face, downloads have surpassed 300,000, and it currently tops OpenRouter's Trending charts. Perhaps most notably, within the influential OpenClaw project (affectionately nicknamed "the crayfish" by developers), Step3.5Flash has quickly risen to become the second most-used model globally. Its combination of speed, stability, and specialized Agent capabilities appears to be winning over programmers worldwide.
A Boost for Open-Source AI
This full-stack open-source approach represents more than just another model release - it's a meaningful contribution to democratizing advanced AI technology. By providing transparent access to both the finished product and its developmental stages, StepZenith is enabling researchers and developers to build better-performing Agents while maintaining control over inference costs.
The move signals growing confidence among domestic AI firms in competing with closed-source alternatives while fostering innovation through community collaboration.
Key Points:
- Complete open-sourcing of Step3.5Flash model including training framework
- 196B parameter sparse MoE architecture with optimized 11B active parameters
- Achieves 350 TPS in coding tasks with strong reasoning capabilities
- Over 300K downloads on Hugging Face, trending on OpenRouter
- Now second most-used model in OpenClaw project globally
