StepFun open-weights Step 5 Preview: 600B/27B MoE, 1M context, SGLang day-0 support
BanghuaZ · x · 2026-10-16
StepFun released Step 5 Preview with open weights (BF16 and FP8 on Hugging Face and ModelScope): a sparse MoE with 600B total / 27B active parameters, a 1M-token context window, and native text+vision input, built for long-horizon agentic work like coding and software engineering.
SGLang offers day-0 inference and RL support, sharing production numbers: on 4×GB300 (TP4) with MTP acceptance length 2.85, FlashInfer TRT-LLM MoE decodes a single request at 299 tokens/s, while MegaMoE reaches 740 output tokens/s per GPU at 64 concurrent requests. The architecture uses hybrid attention (recent window + selected older context), MTP decoding, and cache-aware routing; the Miles framework demos a 79-step GRPO training run.
Related event: StepFun Open-Sources Step 5 Preview: 600B MoE with 1M Context(19 posts)→
More from Infra
- Google's Willow chip cuts quantum errors 40x, raising Q-Day concerns for Bitcoin — jamestagg · 2026-10-16
- Vibe Coding a Data Pipeline Is Easy; Running It at Billions of Rows Is Not — Zachly · 2026-10-16
- Large data centers employ only ~50-150 permanent on-site staff, Microsoft says — brezina · 2026-10-16
- Prediction: Opus 5.5-level models on laptops by 2027, backed by Qwen 27B data — cephaloform · 2026-10-16
- Oracle down 12% in a week, nearly 50% off June high as OpenAI exposure spooks market — alvelda · 2026-10-16
- US data center moratoriums top 400 — an AI bull explains why they're actually good for AI — kevinsxu · 2026-10-16