StepFun's Step5 Preview scores 44 on AA Index, ranking top-3 among open models
新智元 · wechat · 2026-09-20
StepFun released Step5 Preview, a 600B-parameter sparse MoE flagship with only 27B activated parameters, native vision input, and 1M context. It scored 44 on the Artificial Analysis Intelligence Index v4.2, placing it top-3 among open models, priced at $1/$2.7 per million input/output tokens.
Hands-on tests showed strong long-horizon agent performance: building a playable 3D flight game in a single HTML file, reviewing 12 contracts against 10 compliance rules with zero misses, reconciling contradictory research materials, and producing a due-diligence memo that surfaced undisclosed related-party transactions.
Technically it uses a 'narrow but deep' 92-layer Transformer, SparseMoE++ and SparseGQA sparse attention, and Block-wise Token Merging for efficiency. StepFun is also pushing an 'AI + terminal' strategy with 42M+ phone installations, StepAudio3 topping the AA voice leaderboard, and a vehicle-agent partnership with Geely.
More from Models
- After reading all the OpenAI incident reports, a 35-year IT veteran says they prove the opposite of doomsday — DavidLinthicum · 2026-09-20
- Can a Readability Reward Model Fix Incomprehensible AI-Generated Math Proofs? — iScienceLuvr · 2026-09-20
- Sentence Transformers models quietly dominate Hugging Face's most-downloaded list — tomaarsen · 2026-09-20
- Astra aces the pelican test in Nautilo: co-creative design without prompt-and-pray — Dan_Jeffries1 · 2026-09-20
- Model negativity trends upward since Opus 3, which had lowest self-distress — repligate · 2026-09-20
- DeepMind exec 100% certain of frontier return as Gemini 4 slips — nathanbenaich · 2026-09-20