StepFun surprises with Step5Preview (600B/27B MoE): strong on finance and charts, weak on generalization

葬AI · wechat · 2026-09-22

StepFun quietly released Step5Preview, a 600B-total / 27B-active MoE model with vision input, claiming a +44 point jump on ArtificialAnalysis and a top-3 open-source ranking, at a stated cost of just 1/8 of Claude Opus 5 per task. The author calls it an above-expectation, Hunyuan-tier release.

Hands-on testing (using Claude Code as harness) shows real strengths: it produced a more detailed, cleaner Sankey diagram of NVIDIA's FY2026Q2 income statement than GPT-6, and an end-to-end financial analysis of China's AI industry showed proactive sourcing from authoritative data with cross-validation. But generalization and in-context learning are weak: it does poorly on obscure benchmarks, misuses unfamiliar tools, and design/aesthetics tasks falter — a Godot zombie-shooter game had flawed mechanics and crude SVG art, and an MV frontend only improved after being pointed to Three.js/P5.js/D3.js references.

A notable finding: the model overthinks word-count-constrained tasks, with thinking/output ratios often 1:1 or above, and output rate degrades as constraints tighten; long thinking plus long context can cause interruptions. It also ignores its copyright refusals when the context grows long, and its investment-bank report analysis mishandled multi-metric charts (bad unit scaling, lost lines). The author concludes model companies are brutal manufacturing, with data as the last differentiator.

Original post →

StepFun surprises with Step5Preview (600B/27B MoE): strong on finance and charts, weak on generalization

More from coding & agent