Amazon CEO Andy Jassy took to the stage on the company’s re:Invent convention on Tuesday to level off six new textual content material materials, picture, and video experience fashions that it calls Amazon Nova.
This new household of multimodal generative AIs consists of Nova Micro, a text-only mannequin constructed for low-cost, low-latency responses; Nova Lite, a low-cost multimodal mannequin for processing picture, video, and textual content material materials inputs; and Nova Expert, its common objective multimodal mannequin that mixes “accuracy, tempo, and value for a variety of duties,” per the company’s announcement put up. Nova Premier is Amazon’s “most succesful … multimodal fashions for superior reasoning duties,” whereas Nova Canvas is a devoted text-to-image engine and Nova Reel is purpose-built to generate video.
The text-based fashions have been optimized on 15 utterly completely totally different languages. Micro presents a 128,000-token context window whereas each Lite and Expert can address as so much as 300,000 tokens (spherical 225,000 phrases or half-hour of video). The corporate plans to increase the context home house home windows of its higher fashions as so much as 2 million tokens by early subsequent 12 months.
Canvas permits prospects to generate and edit photos utilizing pure language prompts. Reels, which could compete with the likes of Gen-3 Alpha, Kling, and Dall-E 3, can generate clips as so much as six seconds in dimension from each textual content material materials prompts and reference photos. The video generator furthermore presents digital digicam movement administration together with pans and zooms.
“We’ve continued to work on our non-public frontier fashions,” Jassy instructed the assembled crowd, “and people frontier fashions have made an unimaginable quantity of progress over the past 4 to 5 months. And we figured, if now we’ve been discovering value out of them, you’d almost definitely uncover value out of them.”
Jassy furthermore says that these fashions are each among the many many many least dear to function and quickest of their class, although the corporate has nonetheless to position up benchmark information supporting these claims. “We’ve optimized these fashions to work with proprietary methods and APIs, so that you probably can do quite a lot of orchestrated automated steps — agent habits — pretty further merely with these fashions,” he acknowledged. “So I think about these are very compelling.”
The Micro, Lite, and Expert fashions (together with Canvas and Reels) are all at present obtainable to AWS prospects. Premiere is about to realize in Q1 2025.