Step 3.5 Flash: open 11B-active MoE model achieves frontier-level agentic performance

Researchers introduce Step 3.5 Flash, a sparse Mixture-of-Experts model that bridges frontier-level agentic intelligence with computational efficiency by using only 11B active parameters from a 196B foundation.

• Optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce latency and cost of multi-round agentic interactions.
• Utilizes a scalable reinforcement learning framework combining verifiable signals with preference feedback for stable, large-scale off-policy training.
• Achieves 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6, 88.2% on tau2-Bench, 69.0% on BrowseComp, and 51.0% on Terminal-Bench 2.0.
• Performance is comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro.

The model provides a high-density foundation for deploying sophisticated agents in real-world industrial environments by redefining the efficiency frontier.