StepZen Unveils Step 3.7 Flash Large Model, Twice as Fast
China’s large model landscape has reached a new milestone with Step Zenith’s official release and open-sourcing of Step 3.7 Flash. Purpose-built for agent industrialization, this model features deep system-level optimizations for coding, web search, and multimodal workflows.
Architecturally, the model utilizes an advanced sparse Mixture-of-Experts (MoE) structure with 196 billion parameters. This design enables a generation speed of up to 400 tokens per second, drastically cutting latency in high-frequency, multi-turn interactions.

Native Multimodal and Search Enhancement
Step 3.7 Flash offers robust native multimodal comprehension, directly interpreting complex visuals like UI layouts, charts, and documents. This capability allows the model to swiftly transform visual inputs into structured data and generate executable code on the fly.
Furthermore, the model significantly boosts online retrieval and image search functionalities. It actively gathers multi-source evidence from text and images across open networks, ensuring information accuracy through rigorous cross-verification.
Highly Reliable Orchestration and Ecosystem Compatibility
Designed for complex tasks, the model exhibits exceptional stability in long-chain, multi-turn agent workflows. It seamlessly integrates with APIs, browsers, terminals, and Office tools, effectively minimizing execution errors or deviations.
Additionally, the model is deeply optimized for compatibility with mainstream agent development frameworks and tool invocation protocols. This optimization lowers the barrier to entry for developers, streamlining workflow orchestration and ecosystem deployment.
Related article
Anthropic officially launches the new Claude Mythos series Fable5 model
Anthropic, the artificial intelligence firm behind Claude, officially released its latest Fable5 model from the Claude Mythos lineup on Tuesday. This marks the first public debut of this advanced model family, which was previously restricted to inter
Anthropic Claude Teams with NVIDIA GB300 for Azure Launch
Large language models in artificial intelligence are approaching a critical technological milestone. Anthropic has officially made its Claude model series available on Microsoft Azure for enterprise users, marking a significant step beyond basic reso
Microsoft Unveils Scout AI Assistant, Integrating Full Suite with OpenClaw Heritage and Cultivation-Style Customization
Microsoft has advanced its AI agent capabilities with the launch of Scout, a new AI assistant unveiled at the recent Build developer conference. Designed to integrate seamlessly into the Microsoft 365 ecosystem, Scout combines the flexibility of cutt
Related Special Topic Recommendations
Comments (0)
0/500
China’s large model landscape has reached a new milestone with Step Zenith’s official release and open-sourcing of Step 3.7 Flash. Purpose-built for agent industrialization, this model features deep system-level optimizations for coding, web search, and multimodal workflows.
Architecturally, the model utilizes an advanced sparse Mixture-of-Experts (MoE) structure with 196 billion parameters. This design enables a generation speed of up to 400 tokens per second, drastically cutting latency in high-frequency, multi-turn interactions.

Native Multimodal and Search Enhancement
Step 3.7 Flash offers robust native multimodal comprehension, directly interpreting complex visuals like UI layouts, charts, and documents. This capability allows the model to swiftly transform visual inputs into structured data and generate executable code on the fly.
Furthermore, the model significantly boosts online retrieval and image search functionalities. It actively gathers multi-source evidence from text and images across open networks, ensuring information accuracy through rigorous cross-verification.
Highly Reliable Orchestration and Ecosystem Compatibility
Designed for complex tasks, the model exhibits exceptional stability in long-chain, multi-turn agent workflows. It seamlessly integrates with APIs, browsers, terminals, and Office tools, effectively minimizing execution errors or deviations.
Additionally, the model is deeply optimized for compatibility with mainstream agent development frameworks and tool invocation protocols. This optimization lowers the barrier to entry for developers, streamlining workflow orchestration and ecosystem deployment.
Anthropic officially launches the new Claude Mythos series Fable5 model
Anthropic, the artificial intelligence firm behind Claude, officially released its latest Fable5 model from the Claude Mythos lineup on Tuesday. This marks the first public debut of this advanced model family, which was previously restricted to inter
Anthropic Claude Teams with NVIDIA GB300 for Azure Launch
Large language models in artificial intelligence are approaching a critical technological milestone. Anthropic has officially made its Claude model series available on Microsoft Azure for enterprise users, marking a significant step beyond basic reso
Microsoft Unveils Scout AI Assistant, Integrating Full Suite with OpenClaw Heritage and Cultivation-Style Customization
Microsoft has advanced its AI agent capabilities with the launch of Scout, a new AI assistant unveiled at the recent Build developer conference. Designed to integrate seamlessly into the Microsoft 365 ecosystem, Scout combines the flexibility of cutt





Home






