NVIDIA blogs up to 4.45× faster FLUX 3 Image Klein image-to-image on Blackwell B200
Technical blog dated 2026-11-06T22:20:20Z cites TensorRT LLM VisualGen with NVFP4 and fused kernels versus uncompiled BF16 on a single B200 at 2048×2048.
Why it matters
Black Forest Labs released FLUX 3 Image Klein as open weights on October 15. NVIDIA’s November 6 developer blog measures combined optimizations for the Base checkpoint on Blackwell, reporting up to 3.46× faster text-to-image and 4.45× faster image-to-image versus uncompiled BF16. The stack uses TensorRT LLM VisualGen, NVFP4 quantization and fused PrimsTS GEMM kernels. This is an inference-acceleration note, not a new model drop.
Key facts
- NVIDIA developer blog published 2026-11-06T22:20:20Z
- Up to 3.46× T2I and 4.45× I2I vs uncompiled BF16 on one B200 @ 2048×2048
- TensorRT LLM VisualGen + NVFP4 + fused PrimsTS GEMM
- FLUX 3 Image Klein open weights originally 2026-10-15 (BFL)