Add Darwin-36B-Opus VKAE row (11.2x, 25->280.8) + links

Add docker/HF links to Qwen35B leaderboard row

Add JGOS-398B measured row (4.33x, 88->382) to leaderboard series

Remove low-gain Qwen3-32B row (1.1x) from leaderboard

Redact residual method terms (kernel recipe / vLLM/SGLang) to generic serving

Qwen3.5-35B-A3B series -> 601 tok/s (23.4x, peak 10,516); redact method hints

Redact acceleration-method disclosure (graph/vLLM/SDPA) from public rows; keep claim-scoped generic labels

add Qwen3.5-9B(3.41x) Qwen3.6-27B(2.47x) Qwen3.5-35B-A3B MoE(10.35x) measured rows

add Qwen3-8B (1.85x) + Qwen3-32B (1.10x) VKAE CUDA-graph measured rows

remove Gemma 12B/26B-A4B/31B (gemma4_unified, unservable in vLLM, ~1x preview)

remove Gemma E2B entirely (E2B real speedup 1.06x, not a VKAE showcase)

revert E2B: unverified speedup removed (serve logs show ~1x, not 4.1x)

E2B: fill measured B200 VKAE graph numbers (82.9->340.7, 4.1x)

Add VKAE OpenAI compatibility surface

Gate VKAE before-after claims

Update VKAE Gemma model packages and preview benchmark metadata

Add VKAE-Gemma-E2B package status

Fix VKAE claim-scoped benchmark presentation

Mark VKAE candidate confirm16 gate

Update VKAE repeat-2 package benchmark

Productize VKAE-Gemma E4B private preview

Separate optimized VKAE claims from preview measurements

Load benchmark results from latest Hub artifact

Add measured Gemma before-after results

Rename Space to VKAE and add Gemma full-family GPU guide

Add multi-size Gemma measurement matrix

Deploy VKAE-Gemma showcase Space