Accessing Step 3.7 Flash for Free in Hermes Agent

AICodeKinggo watch the original →

the gist

StepFun's new Step 3.7 Flash model is now accessible for free without usage limits via the Hermes Agent portal, providing a high-efficiency option for agentic coding and multimodal tasks.

The Breakthrough

StepFun released Step 3.7 Flash, a sparse mixture-of-experts model optimized for agentic workflows, which is currently available for free with no visible limits through the Hermes Agent portal.

What Actually Worked

  • Accessing the model: Users can authenticate via the Hermes portal by running the hermes model command in their terminal and selecting stepfun/step3.7-flash-free from the available list.
  • Leveraging multimodal agentic workflows: The model utilizes a 1.8 billion parameter vision component to process screenshots, charts, and UI elements, allowing it to perform visual reasoning tasks alongside standard code generation.
  • Utilizing efficient architecture: The model operates with approximately 196 billion total parameters but only 11 billion active parameters, supporting a 256k context window designed for long-running agentic tasks that require tool use and log inspection.
  • Cross-framework compatibility: The model is compatible with multiple agent frameworks, including Hermes Agent, Claude Code, KiloCode, and OpenClaw, and supports MCP-based setups.

Before / After

  • SWE-Bench Pro: Step 3.7 Flash scored 56.3, outperforming Step 3.5 Flash (51.3), Deepseek V4 Flash (55.6), and Gemini 3.5 Flash (55.1).
  • Hermes Agent Benchmarks: Step 3.7 Flash achieved 67.5% accuracy compared to 60% for Step 3.5 Flash.
  • Terminal-Bench 2.1: Step 3.7 Flash scored 59.5, improving upon the 53.4 score of its predecessor.

Context

Step 3.7 Flash is positioned as a high-efficiency model for real-world agentic tasks rather than simple chat interactions. By providing free access through Hermes Agent, the developer enables users to test the model's ability to plan, call tools, and iterate on actual codebases without the friction of API pricing or credit limits. While the model is open-weights under an Apache 2.0 license and can be run locally on hardware with at least 128 GB of unified memory, the Hermes portal remains the most accessible entry point for immediate testing.

Content References

  • #ai
  • #dev-tooling

summary by google/gemini-3.1-flash-lite. probably wrong about something. check the source.