Step 3.5 Flash – Open-source foundation model, supports deep reasoning at speed
static.stepfun.com
February 19, 2026
16 min read
🔥🔥🔥🔥🔥
59/100
Summary
Step 3.5 Flash is an open-source foundation model designed for advanced reasoning and agentic capabilities. Utilizing a sparse Mixture of Experts (MoE) architecture, it activates only 11B of its 196B parameters per token, enabling high intelligence density and real-time interaction.
Key Takeaways
- Step 3.5 Flash is an open-source foundation model with a sparse Mixture of Experts architecture, activating only 11B of its 196B parameters per token for enhanced efficiency and reasoning capabilities.
- The model achieves a generation throughput of 100-300 tokens per second, enabling complex, multi-step reasoning with immediate responsiveness.
- Step 3.5 Flash supports a 256K context window using a 3:1 Sliding Window Attention ratio, significantly reducing computational overhead while maintaining performance on large datasets.
- It is optimized for local deployment on high-end consumer hardware, ensuring data privacy and performance in real-world applications.
Coverage & discussions
Hacker News · static.stepfun.com
Community discussions
Community Sentiment
Positives
- Step 3.5 Flash demonstrates impressive context efficiency, allowing full 256k context streams on a 128GB machine, which enhances its usability for complex tasks.
- The model achieves good inference speeds on Macs, with notable performance metrics like 36 t/s tg and 300 t/s pp, making it practical for real-time applications.
- Its 51% score on Terminal-Bench 2.0 indicates a solid capability for handling sophisticated, long-horizon tasks, which is crucial for advanced AI applications.
- The Mixture of Experts architecture allows selective activation of parameters, optimizing performance while maintaining efficiency, which is a significant advancement in model design.
Concerns
- The model has a tendency to hallucinate significantly, which raises concerns about its reliability for critical applications and necessitates cautious use.
- Some users question the relevance of the 51% score on Terminal-Bench 2.0, suggesting it may not adequately reflect the model's stability in handling complex tasks.
- While the number of parameters is often highlighted, the lack of support for local inference in top models limits their practical applications for many users.
