StepFun Launches Step 5 Preview: 600B MoE Flagship, Open-Weights on Oct 15

StepFun officially released its new flagship foundation model Step 5 Preview on Sept 20, targeting agentic work and professional knowledge work, especially financial analysis. It is a sparse MoE with 600B total / 27B active parameters, native text and vision input, and a 1M context window; weights are scheduled to be open-sourced on Oct 15. The model had earlier been quietly benchmarked by Artificial Analysis, sparking community discussion; the official release confirmed most leaked specs. Reddit users such as nemamdsa12 noted the versioning apparently skipped 4, going from Step 3 straight to 5.

Confirmed

• Architecture: 600B total / 27B active MoE, 1M context, multimodal with native vision (m6, m8, m9, m11, m12, m13, m15, m16, m17)
• Benchmarks: 44 on Artificial Analysis Intelligence Index v4.2, tying Kimi K3 (max) and landing on the cost/intelligence Pareto frontier; ranking wording varies across posts—most say top-3 open-source globally, QbitAI says top-2 (m1, m4, m9, m12, m13, m14, m15, m17); vendor-reported GPQA Diamond 93.5 listed by Benchmark Heaven (m11)
• Pricing: ¥7 input / ¥20 output / ¥0.35 cache per million tokens (about $1/$2.7), roughly one-third to one-fifth of Kimi K3's price; Zhidx estimates single-task cost at 1/8 of Claude Opus 5 (m16, m1, m3, m8, m9, m13, m17)
• Open weights: Oct 15 (m6, m8, m11, m12, m13, m16, m17)
• Capabilities: official demos show long-horizon tasks across software engineering, web apps, 3D workflows and programmable hardware, up to 24 hours of continuous operation, including GPU kernel optimization, with memory of prior results and execution-feedback iteration (m2)
• Finance focus: self-built FinStepBench covering LiveSearch, CorporateValuation and DeepResearch; one case coordinated 950 web fetches in a single agent action, integrating 300k monthly records across 1,000 locations and 25 years into a 17-sheet analysis workbook (m5, m7)
• Ecosystem: trial and docs open, Discord community live, limited-time free Step subscription and Step audio realtime 3 voice model under one API key (m10, m17)

Timeline

• Sept 18: release date listed by Benchmark Heaven (m11)
• Sept 19: Step 5 Preview benchmarked by Artificial Analysis via the StepFun API before any official announcement, sparking discussion (m1, m3, m4, m9)
• Sept 20: official launch with trial, docs and Discord community (m8, m10, m16, m17)
• Oct 15 (planned): weights released (m6, m11, m12, m13)

Why it matters

Step 5 Preview matches Kimi K3's intelligence score at roughly a third of the price, with single-task cost estimated at 1/8 of Claude Opus 5 by Zhidx, pushing open-source capability-per-dollar to a new Pareto frontier. With weights opening on Oct 15, it directly pressures the domestic and global open-model landscape; its 24-hour long-horizon agentic tasks and finance-specific FinStepBench signal a shift of Chinese models from general chat toward deep professional workflows.

2026-09-19 ~ 2026-09-21 · 17 related posts

• [source] StepFun's Step 5 Preview matches Kimi K3 at one-third the price, per Artificial Analysis — No-Head-Royal · 2026-09-19
• Step 5 Preview Quietly Debuts on AA: Score 44, 1M Context, $1/1M Input — teortaxesTex · 2026-09-19
• StepFun's Step 5 Preview Surfaces: 600B Params, Output Priced at 1/5 of Kimi K3 — teortaxesTex · 2026-09-20
• [source] Step 5 Preview Sustains 24-Hour Agent Runs, From GPU Kernels to Post-Training — StepFun_ai · 2026-09-20
• [source] StepFun Launches Step 5 Preview: 600B/27B MoE Flagship, Open Weights Oct 15 — StepFun_ai · 2026-09-20
• Step 5 Preview: 950 Web Fetches and 17-Sheet Workbooks in One Agent Action — StepFun_ai · 2026-09-20
• StepFun's FinStepBench Puts Step 5 Preview Through Financial Scrutiny — StepFun_ai · 2026-09-20
• Step 5 Preview Now Open to Try, Step Plan Subscription Free for Now — StepFun_ai · 2026-09-20
• StepFun's Step5Preview: 600B-param MoE with 27B active, top-3 open-source on AA Index — APPSO · 2026-09-20
• StepFun releases Step5 Preview: 600B/27B MoE, 1M context, beats Opus 5 on kernel optimization — 赛博禅心 · 2026-09-20
• StepFun's Step5 Preview scores 44 on AA Index, ranking top-3 among open models — 新智元 · 2026-09-20
• StepFun Launches Step 5 Preview: 600B MoE with 27B Active, Open Weights on Oct 15 — airesearch12 · 2026-09-20
• StepFun's Step-5 Preview Leaks: 600B/27B Active, 1M Context, GPQA Diamond 93.5 (Vendor Numbers) — airesearch12 · 2026-09-20
• StepFun launches Step5 Preview: 600B MoE with 27B active params, claims 1/8 the task cost of Claude Opus 5 — 智东西 · 2026-09-20
• StepFun previews Step 5: 600B MoE with 27B active at $1/$2.7 per M tokens, open weights next month — nemamdsa12 · 2026-09-21
• StepFun launches Step 5 Preview: 600B MoE, 27B active, 1M context at 65% lower cost — realmrfakename · 2026-09-21
• Step5 Preview launches with 600B params, 27B active, 1M context — top-2 open model at 12.5% of Opus 5 cost — 量子位 · 2026-09-21