StepFun drops Step 5 Preview 600B MoE
StepFun introduced Step 5 Preview, a 600-billion-parameter Mixture-of-Experts (MoE) model that routes through 27 billion active parameters per token. Equipped with a 1-million-token context window and native multimodal vision support, the model is targeted specifically at high-demand domains including software development and financial analytics. StepFun claims that Step 5 Preview matches the frontier performance benchmarks of GLM 5.3 and Kimi K3 while cutting per-task inference costs by roughly 65%, with open model weights scheduled for public release on October 15.
StepFun is proving that sparse MoE architectures remain the fastest path to undercutting closed-source frontier economics without sacrificing reasoning depth.
- –Sparse efficiency: Activating only 27B out of 600B parameters delivers high-capacity modeling with manageable inference latency and hardware requirements.
- –Enterprise focus: Tailoring multimodal capabilities and a 1M token context specifically to finance and coding tackles two of the most compute-heavy and commercially viable sectors.
- –Open-weight timing: Committing to release weights on October 15 will put direct pressure on API-only rivals across Asian and global markets.
- –Margin compression: Slashing costs by 65% relative to comparable frontier models accelerates price deflation for agentic workflows and long-context processing.
DISCOVERED
1h ago
2026-09-21
PUBLISHED
1h ago
2026-09-21
RELEVANCE
AUTHOR
WorldofAI