StepFun's Step5Preview: 600B-param MoE with 27B active, top-3 open-source on AA Index
APPSO · wechat · 2026-09-20
StepFun released Step5Preview, a new open-source flagship that scores 44 on the Artificial Analysis Intelligence Index, placing it top-3 among open-source models and on the capability-cost Pareto frontier.
- Architecture: sparse MoE with 600B total / 27B active parameters, native text+vision input, 1M context, aimed at coding, software engineering, finance and other long-context, multi-turn tool-use workloads.
- APPSO hands-on: strong Three.js game/scene generation (gondola ride, hot-air balloons, Niagara Falls), a web OS judged more complete than DeepSeek V4 attempts, photo-to-3D reconstruction in minutes; on the FrontierFinance benchmark (220 expert questions, 11,543 criteria) results approach Claude Opus 5.
- Omni-modal & devices: StepEdge on-device model, StepAudio3 Realtime scoring 98.9%/99.7% on voice tests, ASR WER of 1.7% tied for global first; 42M+ phone installations serving 20M daily, in-vehicle agent Super Eva on Zeekr 8X.
- Ambition: STEPX device brand and STEPX Neo agent phone, betting on Personal Agents as the next wave after Coding Agents.
Related event: StepFun Launches Step 5 Preview, Open-Weights on Oct 15(6 posts)→
More from Models
- 100-run trolley problem test: Jev pulls the lever 99% of the time — FinanceYF5 · 2026-09-20
- Parallel-sampling decision model Jev routes in ~1s vs 4-14s for regular LLMs — TigerOk4538 · 2026-09-20
- Working memory seems to scale with model size, according to one researcher's observations — repligate · 2026-09-20
- Burkov disputes Stanford AI Index: does the US really lead with 59 notable models? — burkov · 2026-09-20
- JEV watch: a model that returns typed probability decisions instead of text — sven_ai · 2026-09-20
- Open questions on JEV: model size, calibration reliability, and a future fine-tuning API — vykthur · 2026-09-20