MiniMax dropped H3 (Hailuo 3.0) on July 31 and says the weights land within days. That’s the headline: China’s open-source wave, which already flooded text and image models, just reached video generation for the first time. H3 is a text/image/video/audio-to-video model that outputs native 2K clips at 24fps, 5–15 seconds long, with stereo audio —… Continue reading

Black Forest Labs, the team behind the FLUX image models, just shipped FLUX 3 — and it’s no longer just an image generator. It’s one model that produces image, video, and audio from a single set of weights, jointly trained instead of three separate models bolted together behind one API. What it actually does Text-to-video,… Continue reading

Alibaba’s Qwen team shipped the third generation of its image model today. HN front page within hours — 74 points, 40 comments before lunch. What it actually is One foundation model that generates and edits, no separate editing checkpoint. It takes instructions up to 4.5k tokens (2.0 capped around 1k), renders text down to 10px,… Continue reading

Every AI image generator hands you a flat JPEG. You want to move the logo? Too bad, it’s baked into the pixels. ByteDance’s new flagship image model, launched July 8, kills that problem. What Seedream 5.0 Pro actually does It’s a text-to-image model with one killer trick: layered output. One render splits into a background… Continue reading

Meta killed off Emu. The first media generation models from Superintelligence Labs — the org Alexandr Wang runs — are called Muse Image and Muse Video, and they’re nothing like the Llama-era tools they replace. Image generation that acts like an agent Muse Image isn’t plain prompt-to-image. It works agentically: it calls search and code… Continue reading

Google shipped Google Gemini Omni Flash on June 30 — the cheap, fast tier of its Omni family, built for one thing: making video by talking to it. Describe a scene in plain English, get a clip. Then keep talking to fix it: “darken the sky,” “add a dog,” “make it slower.” No timeline, no… Continue reading

Most AI video models top out at 5-10 seconds, then fake longer clips by gluing segments together — and the seams show. ByteDance’s Seedance 2.5, unveiled at the Volcano Engine FORCE conference, generates a single native 30-second clip in one pass. Character faces, lighting, and motion hold steady the whole way through, because audio and… Continue reading

Google DeepMind just shipped the cheapest, fastest member of the Nano Banana family. Model ID: gemini-3.1-flash-lite-image. It’s an image-generation model built for one thing — throughput. One image in under 4 seconds, roughly 2.7× faster than Gemini 3.1 Flash Image, at a flat $0.034 per 1,000 images. On Artificial Analysis’s text-to-image leaderboard it sits at… Continue reading