Summary
Step 3.7 Flash is a 198 billion parameter sparse mixture-of-experts model from Step One, activating only 11 billion parameters per token during inference. It ships with a 256K context window, native image and video understanding, three reasoning levels (low, medium, high), Apache 2.0 licensing, and support in both vLLM and SGLang. Because running it locally requires multiple high-end GPUs, Fahd Mirza tests it via API using the Hermes agentic framework.
The agentic evaluation is the highlight: asked to analyze a Python search monitoring app and generate a structured markdown report, the model reads five files, grasps the full architecture, and produces a detailed report in a single shot — using parallel file reads, terminal commands, and file writes without hand-holding. The multilingual test covers 80+ languages including obscure regional ones like Gutnish, Faroese, and Tigrinya, with correct rendering of RTL scripts such as Arabic, Urdu, and Burmese. A multimodal task then asks the model to analyze a lightning storm image and generate a working index.html visual based on it — which it completes through Hermes tooling.
On benchmarks, Step 3.7 Flash reportedly leads on the Claude eval, which tests instruction-following across long multi-turn tool-use sessions, outperforming both Gemini 2.5 Pro and GPT-4. For developers evaluating open-weight models for agentic coding pipelines, this video offers a practical real-world comparison against the kinds of tasks where model reliability actually matters.
📺 Source: Fahd Mirza · Published May 29, 2026
🏷️ Format: Hands On Build
