Pika released H3 Flex, a post-trained MiniMax H3 model family that generates video with synchronized audio in six sampling steps. Public weights, code and documented paths cover text-to-video, first- and last-frame generation, and reference-conditioned video at native 480p or 768p with up to 12 image, video or audio references.
Why it matters: The release gives video teams an inspectable local path for rapid multimodal iteration instead of limiting the workflow to a hosted creator. The useful boundary is narrower than the launch headline: local artifacts are adapters derived from MiniMax H3, while…
Black Forest Labs released FLUX 3 Image Klein as 5.6B open weights for generation and editing. NVIDIA now provides TensorRT LLM VisualGen configurations and reports that compilation, NVFP4, sparse attention and caching cut single-B200 2048×2048 latency by as much as 3.46× for text-to-image and 4.45× for image-to-image versus uncompiled BF16.
Why it matters: The added configurations turn a model announcement into a reproducible serving path with measured speed-quality tradeoffs. Operators can compare quantization, sparse attention, caching and multi-GPU scaling rather than treating a single headline latency as…
Nunchaku Lite checkpoints now load through standard Diffusers pipelines, with Hub-delivered kernels and no separate inference engine. The release includes prequantized models plus a toolkit and tutorial for quantizing additional diffusion transformers.
Why it matters: A standard loading path makes low-memory image inference easier to adopt, test and combine with existing schedulers, LoRAs and offloading tools.