Step Releases 3.7 Flash, Open-Source Multimodal Model Built for Agent Efficiency
Key Takeaways
- ▸Step 3.7 Flash achieves +5% on SWE-Bench Pro and 6.1% on Terminal-Bench 2.1 versus Step 3.5 Flash, with substantially improved balance across diverse agent harnesses
- ▸Native multimodal understanding, enhanced web/visual search, and reliable tool orchestration enable Step 3.7 Flash to handle complex agentic workflows with lower failure rates
- ▸Advisor Mode reaches 97% of Claude Opus 4.6's coding performance while maintaining Flash-tier efficiency, addressing the cost-quality tradeoff for production agents
Summary
Step has released Step 3.7 Flash, an open-source multimodal model explicitly designed for agent workloads and code generation. The model prioritizes operational efficiency while maintaining strong performance across diverse agent harnesses, marking a significant step toward production-grade agentic AI systems. Available on GitHub, HuggingFace, and ModelScope, Step 3.7 Flash achieves a +5% improvement on SWE-Bench Pro and 6.1% on Terminal-Bench 2.1 compared to its predecessor, with notably more balanced performance across different agent frameworks including Claude Code, KiloCode, Hermes Agent, and OpenClaw.
The model's core strengths lie in native multimodal understanding—handling product UIs, documents, charts, and natural scenes—alongside enhanced web and visual search capabilities and reliable tool orchestration that minimizes drift and failed tool calls. A standout feature is Advisor Mode, a hybrid approach where Step 3.7 Flash handles most agentic tasks autonomously while consulting a larger frontier model only at critical decision points, achieving 97% of Claude Opus 4.6's coding performance while maintaining Flash-tier cost efficiency.
As an open-source release, Step 3.7 Flash directly addresses the emerging gap between closed-source frontier models and practical, deployable alternatives. The full model weights and broad ecosystem compatibility reduce integration friction for teams already invested in popular agent platforms, making efficient, multimodal agents immediately accessible to developers and enterprises.
- Open-source availability on GitHub, HuggingFace, and ModelScope with compatibility across Claude Code, KiloCode, Hermes Agent, and OpenClaw reduces deployment barriers
