SGLang reproduces large-scale EP on 96 H100s

Reproduced large-scale EP on 96 H100s — $0.20 per 1 million output tokens

Why it matters

An open-source reproduction of frontier-class MoE serving

Source — Primary sources first; secondary reports are labelled “Press”

Officiallmsys.orghttps://lmsys.org/blog/2025-05-05-large-scale-ep/

Added in Week 35, 2026 · See that week

Same week

How far can reasoning models scale?OfficialUS Copyright Office releases pre-publication Copyright and AI Part 3GovernmentPyTorch Foundation takes in vLLM as a hosted projectOfficial

Nearby on the same axis

PyTorch Foundation takes in vLLM as a hosted projectOfficialGemma 3 QAT int4OfficialNVIDIA announces Dynamo at GTCOfficialNVIDIA introduces NVFP4Official