Claude Code Meets Step‑3.7‑Flash: Small Model, Big Multimodal Power

Step‑3.7‑Flash, released by the Chinese model company 阶跃, is positioned as a "high‑efficiency flash model for production‑grade agents." Compared with earlier flash models, it adds native multimodal support while keeping the same 196 B total parameters and 11 B activation footprint, plus a 1.8 B visual encoder, 256 K context window, and a claimed peak generation speed of 400 TPS. The pricing (input cache ¥0.27, miss ¥1.35, output ¥8.1) follows the domestic trend of low‑cost, high‑frequency usage.

The model integrates vision directly into the agent workflow: images can be dropped into the dialogue and used without a separate visual model. Built‑in web search and visual retrieval become part of the reasoning chain, allowing the model to answer visual questions (SimpleVQA 79.16%) and perform fine‑grained visual tasks (V* benchmark 95.29%). Coding benchmarks show SWE‑Bench Pro 56.26% and Terminal‑Bench ~60%.

Using the author’s lightweight routing tool ccmr, Step‑3.7‑Flash is added to the Claude Code model list via the /model step-3.7-flash endpoint. The first hands‑on task feeds a hand‑drawn sketch of a notebook‑style web UI to the model, prompting it to generate a single‑page website. The output reproduces the sketch 1:1, then a second iteration refines the design with better fonts, colors, and layout, demonstrating fast generation and acceptable quality.

The second task evaluates visual understanding on a UNet++ diagram from medical image segmentation. The model first inspects the figure, then performs two rounds of web search to verify the method, correctly identifying the architecture (Nested U‑Net, Zhou et al. 2018) and explaining its components (down‑sampling levels, skip‑connection convolutions, dense jumps, deep supervision). This showcases the model’s ability to combine visual parsing with external verification.

The third task challenges the model with a PointRend diagram from CVPR 2020. Initially, the model describes the pipeline (CNN backbone, coarse prediction, point sampling, MLP refinement) and launches multiple search rounds, retrieving original papers and confirming the method. It briefly misinterprets a silhouette as a pose‑estimation task but self‑corrects after further retrieval, illustrating the built‑in search‑and‑verify loop.

Overall, the experiments suggest that flash models like Step‑3.7‑Flash are not meant to replace flagship LLMs but to handle high‑frequency, multi‑step agent workloads efficiently. Their value lies in completing entire task chains reliably and cheaply, a shift from peak intelligence metrics to end‑to‑end task efficiency.