Google shipped Gemma 4 today. A few hours later we had the 12B running on an NVIDIA GB10 (Grace-Blackwell, 121GB unified memory). Day-one, fully local, GPU-offloaded.
Honest field notes — real numbers, real caveats. 🧵
96GB of VRAM running Mixtral, Qwen2, and DeepSeek locally — finally put GPT-4 performance claims to the test with real token/sec benchmarks that'll shock you.