The vLLM team made Kimi K3 run 57% faster on Google TPUs
Inferact, founded by the creators of vLLM, released a megakernel for TPU v7: 16 Google chips beat 16 NVIDIA GB200 GPUs by 57% on Kimi K3 inference. The code is open source.
Studio scenes: how to stitch one video from many scenes
Add whole scenes — a talking avatar, an image or a clip — and the section stitches them into one video at the resolution and aspect ratio you choose.
Ideogram API and MCP: how to add image generation to your
Ideogram V3: capabilities, the /v1/images/generations endpoint, MCP tools and getting a key.
HyperFrames: rendering an HTML animation into video
Upload a composition as a zip or by link — the section renders frames and animation into a finished video at the resolution and format you pick. We cover what to set, how to launch a render and what it costs.
Animate a photo with AI: from a still image to motion
One still photo becomes a short moving clip: which NeuralSpace models to use, how to describe motion in words and what the price depends on.