A rhythm completed in three steps

Released on September 20 as a closed service; topped OpenRouter's Trending chart on October 10; open-sourced in full on October 15 — in 25 days, StepFun completed a full cycle of the new "validate first, then open-source" cadence for Chinese frontier models. Per multi-source reports (Beijing News, Shangguan News, The Paper, IThome), Step 5 Preview's BF16 and FP8 weights, technical report, open-source tooling and deployment guides are live on Hugging Face and ModelScope for free global download.

Model file (cross-verified)

· Architecture: sparse MoE, 600B total parameters, ~27B active per token

· Context: 1M-token window, natively multimodal (text + vision input)

· Design: 92-layer narrow-deep stack, optimized for long-horizon agent tasks

· Focus: AI coding, software engineering, professional knowledge work, finance

· Ranking: #2 on OpenRouter's weekly and daily token-usage charts

· Ecosystem: integrated into Tencent WorkBuddy, free trial through October 25

The most telling detail: five Chinese chip vendors adapt on day one

The adaptation camp assembled the same day the weights dropped: Huawei Ascend, MetaX and Biren announced 0-day support; Moore Threads declared Day-0 adaptation on its MTT S5000 cards with the MUSA software stack; Hygon's DCU completed adaptation the same day. This is not ceremonial — in the very same week, Chinese GPU makers were racing rack-scale supernodes at CIIF (see our companion story). "Chinese model × Chinese silicon" has, for the first time, formed a mutually reinforcing combo.

The open-source release day is also the delivery day for domestic compute: for the first time, model and chips shipped on the same date.

Open source is not the finish line — it is the funnel

The strategy is legible: validate real usage as a closed service first — the #2 spot on OpenRouter's weekly chart is the best advertisement — then trade full weights for ecosystem position. The FP8 release reserves cost headroom for inference, and the 600B-total / 27B-active sparsity clearly accommodates the memory reality of domestic accelerators.

In 2026, when open flagship models have become table stakes, long-horizon agents and million-token contexts are the daily engineering reality for NineZenith's Tianxing agent platform: models change, chips change; the invariant delivery standard is "finish the task reliably". Whoever turns open weights into closed business loops truly wins this round of open source.

Trend watch: as release cadence compresses from years to months, competition shifts from the model itself to the adaptation matrix, inference economics and scenario engineering. (Information synthesized from public reports by Beijing News, Shangguan News, The Paper, IThome and Guandian)