StepFun releases Step5 Preview: 600B/27B MoE, 1M context, beats Opus 5 on kernel optimization
赛博禅心 · wechat · 2026-09-20
StepFun launched flagship Step5 Preview: 600B total / 27B active params, 1M context, native text+vision. Priced at ¥7 input / ¥20 output / ¥0.35 cache per M tokens; open-sourcing Oct 15. Scores 44 on Artificial Analysis, top-3 among open models, free on AGIBar for a week.
Highlights
- Positioned for software engineering, deep research, and long-horizon agent workflows
- Scores 49.0% avg@4 on its own StepCodeBench (553 repos, 33 languages); low/mid-difficulty tasks near Claude Opus 5
- Kernel optimization: 508 TFLOPS on one H100 in 24h, beating Claude Opus 5 (493) and Kimi K3 (307)
- Self-improvement experiment: lifted Qwen3-30B-A3B's AIME24 accuracy from 53.3% to 60% in 24h
- Played Pokémon Red unoptimized: 3,093 turns, 6.29M tokens, 3 badges
Tech: 92-layer "narrow but deep" design, SparseGQA + block-wise token merging cutting indexer cost to 1/8, bit-wise on-policy RL alignment with <10% overhead, and model-in-the-loop data generation with anti-hacking audits.
Related event: StepFun Launches Step 5 Preview: 600B MoE Flagship, Open-Weights on Oct 15(17 posts)→
More from coding & agent
- SemiAnalysis: Persistent agents now eat 45% of our AI spend, and Perplexity's harness beats Claude Tags — dylan522p · 2026-11-12
- Ora's Ranker 3.0 grades whether agents can actually use your site; scores drop 40+ points — yenkel · 2026-11-12
- Gemini API Managed Agents now support self-hosted sandboxes, incl. Vercel — steren · 2026-11-12
- Rayline launches ARC, a per-turn model router claiming 90% cost cuts — testingcatalog · 2026-11-12
- Karpathy's llm-wiki: a pattern for building personal knowledge bases with LLM agents — stochasticchasm · 2026-11-12
- Agent self-improvement with frozen weights: failure traces add just 4pp, blind search gets 60% — burny_tech · 2026-11-12