gemma-4
🤖FFPA: Extends FlashAttention-2 via Split-D for large headdims, 1.5x~3×↑🎉 vs SDPA, up to 430T🎉 on H200.
MCP server that gives vision to any text-only LLM — 100% local, no API bills. Gemma 4 E2B (LiteRT) for vision + SDXL-Turbo for image gen