Amazon New AI Models 'NOVA' Stun The Entire Industry!

So I'm excited to share and announce the launch of Amazon Nova, which are our new state of the art foundation models that deliver frontier intelligence and industry leading price performance.

You can see here that Amazon literally just announced their frontier models. It's really fascinating to see other companies get in on the AI space. So watch this entire video because it's going to show you guys everything you need to know about Amazon's new frontier models and their capabilities.

So in this intelligence set of models there are four flavors. The first is Micro, which is a text only model, which means you feed it text and outputs text. It's laser fast, very cost effective and our internal builders are really enjoying it for a lot of their simple tasks. And then we have three flavors of multimodal models. And so multimodal models, you imp, you can input text or image or video and you output text. And so each of these are in ascending order of size and intelligence. The Micro, Light and Pro models are generally available today. The premier model will be available in the Q1 timeframe. So I'm gonna share a few benchmarks. I'll just say that we used external published benchmarks whenever we could. When they weren't available, we did it ourselves. We published the methodology on our website. So you can try and replicate if you like. So I'll just share some of the benchmarks. So on the micro model you can see it is a very competitive model. If you look at the raw numbers relative to the leading models in this class, Llama and Google's Gemini, on the raw numbers, it benchmarks better on all the variables versus llama and 12 or 13 versus Gemini. But if you do statistical significance testing, which we did, we just took all the numbers that were overlapping in the 90, 95 confidence interval and we called those equal. So if you look at that way, which I will moving forward, you can see that we are equal or better on all the benchmarks compared to Llama and Gemini in these, in this class of models. If I look at the light model, it's a very similar story, very competitive. If you compare novo light to OpenAI's GPT4O mini, you can see that we're equal or better on 17 of the 19 benchmarks, equal or better on 17 of the 21 benchmarks versus Gemini, and then equal or better on Haiku 3.5 on 10 of the 12 benchmarks. Haiku isn't doing images or video yet, so we couldn't benchmark on as many dimensions. But again, a very competitive model. And then if you look at Pro Same story if you Compare it to GPT4O, it's equal or better on 17 of the 20 benchmarks. It's equal or better on 16 of the 21 benchmarks versus Gemini. The very best model in this class of models is Sonnet v2.3.5. But even here you can see that our PRO model is equal or better on about half of those. And on the ones we're not, it's very competitive. And you're going to like the cost and the latency characteristics here. And then our premier model, which will be our largest multimodal model, will be available in the Q1 time frame now. So that's four very competitive, compelling intelligence models. But there are some other things that I think you're going to really like about these models. First, they're really cost effective. They're about 75% less expensive than the other leading models in Bedrock. Two, they are fast. They're the fastest models that you'll see with respect to latency. We'll also make the Nova models available in the SKU that Peter was talking about last night in the latency optimized inference SKU as well. Very fast. And then they're not just integrated, they're just not in Bedrock, but they're integrated deeply with all the features in Bedrock that any model provider can use. It's just that this team took the time to do them. So that means that you get fine tuning. Increasingly, a lot of our app builders for generative AI, they want to do fine tuning with labeled examples to make the applications perform better. You'll also be the Nova models are also integrated with the distillation feature the mattress talked about, so you can infuse the intelligence of bigger models into smaller models that are more cost effective and lower latency. It's deeply integrated in Bedrock's knowledge basis so that you can use RAG to ground your answers in your own data. And then also we have optimized these models to work with the proprietary systems and APIs so that you can actually do multiple orchestrated, automatic steps, agentic behavior much more easily with these models. So I think these are very compelling. I'm looking forward to taking a shot at them and using them now. Customers want to actually do more around generative AI than just with outputs that are text. They also have a lot of needs around images, around video, and there's lots of examples of it, but simple ones are advertising or marketing or trading materials. And so we've worked hard. You know, it's expensive. There aren't a lot of Options out here, they're not easy to do yourself. We've worked hard on this problem. I'm excited to announce two more models for you. First is Amazon Nova Canvas, which is our state of the art image generation model. And so canvas allows you to input natural language text, get images back. They're beautiful images, they're studio quality images. It allows you to edit images using natural language or text inputs. It gives you controls for color scheme and layout. It has a number of built in controls for responsible use of AI, including watermarking for traceability as well as content moderation to limit the generation of harmful content. And we, we benchmark this as well. We tried to benchmark it versus some of the other state of the art players in this space. In this case we picked, you know, typically what people consider the two leaders here, which are, you'll see Dolly 3 and stable diffusion 3.5. And we benchmarked on the two variables that matter most, which is image quality and instruction following. And you can see that Canvas outperforms both of them on both those dimensions. We also did a human evaluation where you saw similar types of results. So this is a compelling model. And then of course we also want to allow you to have it be easy to generate video. And so we're excited to announce the launch of Amazon Nova Real, which is our state of the art video generation model. So again with Real, it's studio quality video. It's really stunning videos that you can create. It gives you full control of the camera, lets you have motion control, lets you do panning. It lets you do 360 degree rotation and zoom. It also has built in AI controls for safe AI including watermarking and content moderation. We'll launch it with the ability to do six second videos, which works really well for a lot of marketing and advertising. On its way up to two minute videos in the next few months, we benchmark this as well. There aren't really many video generation services that have an API and then none of them have automated benchmarks. So we just took one of the we benchmark with human evaluation versus one of the leaders here in Runway and you're going to see again that real benchmarks very favorably relative to others.

So.

That six new frontier models for you. What's going to be next for us in Nova? Well, the first thing is the team is going to be working really hard over the next year on the second generation of these models. But I also have a couple things that I thought I'd give you a sneak peek into. The first is that in the Q1 timeframe, we anticipate giving you a speech to speech model which will allow you to input speech and get speech back very fluent, very fast. And then around mid here, we're going to give you an any to any model. So this is really multimodal to multimodal. So you'll be able to input text, speech, images or video and output text, speech, images and video. This is the future of how frontier models are going to be built and consumed. And we're really looking forward to giving this to you. So you may be asking yourself, how should I think about AWS's model strategy? They have very deep partnerships with a number of model providers. They have some of their own models now. And the way I would tell you to think about it is the way that we always provide you selection in everything we do, which is that we are going to give you the broadest and best functionality you can find anywhere. And what that's going to mean is it's going to mean choice. The reality is that all of you are going to use different models for different reasons at different times, which by the way, is the way the real world works. Human beings don't go to one human being for expertise. In every single area, you have different human beings who are great at different things. You're going to sometimes be optimizing for coding, sometimes for math, sometimes for integration with rag, sometimes for agentic needs, sometimes for lower latency, sometimes for cost, most of the time for some combination of these. And at aws, we are going to give you the very best combination of all of these, as we always do. And we think we've added some pretty interesting models to the mix today. But the great thing is, is that all of these models are available for you in Bedrock, and you can use them in whatever combination you want. And you can experiment, you can change over time. And we will give you that selection and that choice today as well as in the future. So with that, I will say have at it. Giddy up and back to Matt.

So we could clearly see here that Amazon has introduced its own frontier models that are performing really, really well. And we actually take a look at the wider AI market by taking a look at this artificial analysis quality index. We can see that they actually do perform at a decent rate. They are firmly wedged within the AI ecosphere as a model that many individuals are going to use. Of course, it's probably going to be a lot of Amazon AWS customers, but nonetheless, it is surprising that a company can just come out with their own frontier models that are on par with some of the leading AI companies. So it'll be interesting to see what Amazon holds for us in the future as they develop the AI ecosystem even further.