vLLM Ships Day-0 Support for Step-5 Preview: 604B MoE with 1M Context and RL Recipes
vllm_project · x · 2026-10-15
vLLM announced full support for StepFun's Step-5 Preview: a multimodal MoE with hybrid sparse/sliding-window attention, 1M-token context, and MTP. Highlights: official Docker image, FP8 on 4×B300 (TP4/TEP4/DEP4 with MegaMoE on Blackwell, 8×H200 otherwise), an NVFP4 checkpoint, a P/D prototype on 12×GB300 hitting 91.24% prefix-cache hit rate with zero errors across two 1,000-request OCR runs, and Day-0 RL support via verl (GRPO + LoRA on FSDP2, vLLM rollouts, R3 routing replay, adapter-only weight sync).
Related event: vLLM Ships Day-0 Support for Step-5 Preview with Sparse Attention and MoE(4 posts)→
More from Infra
- When Claude refuses your insane GPU optimization, DeepSeek gets the call — mike64_t · 2026-10-15
- KAIST's KV4-RL: INT4 KV cache quantization speeds RLVR rollouts 1.8x without accuracy loss — kaist-ai · 2026-10-15
- Nordics race to deploy 1GW of AI compute, with a 1,152-GPU DGX GB300 supercomputer — NVIDIA Blog · 2026-10-15
- ASML raises 2026 forecast to €45B on 45% EUV demand surge but warns of tool shortages through 2028 — CurieuxExplorer · 2026-10-15
- NVIDIA to showcase neural rendering, physical AI and robotics at SIGGRAPH Asia 2026 — tomaszbednarz · 2026-10-15
- Reddit user builds 256GB VRAM local AI rig with 4x CMP 170HX cards in a chicken-coop shed — sevendaysworth · 2026-10-15