StepFun open-sources Step Code v0.1.4, details 92-layer agent-first Step 5 architecture
StepFun_ai · x · 2026-10-15
StepFun updated Step Code to v0.1.4, its MIT-licensed open-source agent harness built around Step 5 Preview (900 GitHub stars), the fourth release since open-sourcing, driven mostly by community feedback.
The framework helps agents track goals on long, complex tasks, keep making progress, and verify against checkable results whether the job is done.
The accompanying technical blog details the architecture/training co-design for agentic work:
- 92-layer narrow-and-deep architecture for implicit multi-hop reasoning
- Sparse GQA to cut million-token context cost, with 8-token block indexing reducing indexing and top-k overhead to 1/8
- FP32 precision on key components
More from coding & agent
- NousResearch demos Hermes agent building HTML5 apps and iOS apps in few shots, open-sourcing weights and tech report — kuchaev · 2026-10-15
- 220-question test: 70% of model web searches were unnecessary, cost 75x more tokens — demian_ai · 2026-10-15
- 'The safety cult is tokenmaxxing': agents now monitoring agents — zsakib_ · 2026-10-15
- Free Future of AI Search conference: late interaction, MUVERA benchmarks and agentic search — IgorCarron · 2026-10-15
- Zero-dependency prayer-time MCP server brings reminders to Claude, Cursor and ChatGPT — TareqN · 2026-10-15
- Composable models: routing, memory, and verification models will change how apps are built — iamrobotbear · 2026-10-15