00:00Revenue was up 283 percentage. Can you sustain that? Were you happy with those
00:07results? Yeah, I think we continued the growth momentum. As you said, our
00:14revenue in the first half year already grew more than 280 percent. It's already
00:19became a 1.5 times of last year's a whole year revenue. And up to now, our AR
00:25already surpassed 800 million USD, growing up more than four times within only
00:32six months. And all these numbers are driven by our model update. So in the
00:37first half year, we update several versions of our language model and
00:41multi-modality models. And the token consumption in July already became 20
00:45times of the numbers in January. Okay. Do you have a full-year target on ARR?
00:52I think we shared the number before. It's like a 1 billion USD ARR at the end of
00:57the year. We have more confidence on that. Okay. Now, you mentioned the fact
01:03that revenue is growing very quickly because of all the updates you're doing
01:08to your models. Is the relationship between how quickly you update the
01:12models and you release new models, is that very clear in as far as the revenue
01:16impact is concerned? In other words, what I'm trying to get to is, are you looking
01:19to increase the cadence of your releases more frequently so it drives business?
01:24I think the key of our company is in our name, Minimax. Minimizing the cost and
01:30maximizing the intelligence. So we think the intelligence and the cost efficiency is
01:35not a trade-off. It's even have more synergy. So if you have more and more
01:40cost-efficient models and inference-efficient model, it's all only benefit for
01:45your inference, for your revenue, but also benefit your all-training process. Because
01:51including generating synthetic data, analyzing data, doing evaluation and post-training
01:59RL, all this part will need to inference compute. So the cost efficiency definitely
02:04also accelerates your model intelligence level, which will definitely drive to a larger
02:10total consumption of your model, drives to the commercialization success of the company.
02:18Okay. So obviously it focuses on the product and making that better. What about the losses
02:25though? And I know you're still in the growth phase, so you might not be too focused on your
02:30reporting revenue growth, but the losses seem to be deepening. When do you expect your losses
02:34to peak? So as I mentioned, the revenue in the first half year already grows more than 280%. Yes. And
02:43our R&D spending is only growing around like 130%. So revenue grows much faster than our R&D spending.
02:53So as I said, our key strategy is deliver both frontier models with the price
03:00competitiveness. That will drive the company's success. And how would you consider right now,
03:05when you look at the competitive landscape, how worried are you about what they are launching,
03:10as opposed to what you guys are doing? How often do you follow the competition? And who do you
03:15consider their biggest competition? They are, I think all they did a great job actually. So we have our
03:21strategy. First, I could be studying our name, Minimax. And the second is openness. We think
03:27openness is our assets. I think we open source, open with all our frontier models. And for example,
03:35we just released the H3 three weeks ago. It was downloaded more than 24 million times in Hagenfils
03:42and have more than 300 derivatives. It already became the most downloaded apps, models in the world
03:49currently. And openness is an asset for us. The feedbacks coming from open communities became a flywheel for our
03:57future R&D, which close to the source labs cannot match. At the same time... Why is that? If you
04:02could clarify,
04:03why do you think they can't match that? Because all the open communities can help us to generate more
04:08SKUs. For example, our partners, Phil, FAL, they designed a model based on our foundation models that
04:18they cut off some part of the inference and make a model called Minimax H3 Max. This model can generate
04:25five seconds videos within only three seconds. It unlock a new market, which is like 24 seven AI live
04:34streams. So working with our partners under our commercial license, we unlock the create even larger
04:42market for all of the world. Yeah. And I think your stock actually did very well on the AI live
04:47stream that
04:47you just alluded to just now very, very recently. Now, obviously, you've raised a lot of money. You've
04:52raised a lot of money in Hong Kong at the start of this year. I'm looking at your balance sheet.
04:57You
04:57haven't actually used a lot of that money you've raised. You're still sitting on quite a bit of cash.
05:01Where do you see you using that cash over the next few months? I think pushing the model to the
05:07intelligent limit is definitely always the key because we can build up a stable infrastructure,
05:13build a really great foundation models and also provide no matter is the enterprise APIs or harness
05:21for the end users to use the other models. And also talent is really important for this industry.
05:28We are acquiring really great people around the world. All these are important for foundation model
05:33companies. But as I said, as I said, the revenue grows much faster than cost. So in the future,
05:38you will see we will have a more and more healthy business model. And just generally speaking,
05:44how do you differentiate yourself from rivals on the pricing side in terms of your pricing strategy?
05:49I know that's different tiers, but how are you? What's your philosophy? So deliver a frontier
05:55models with a price competitiveness is our key strategy. Does that mean cheap?
06:02It's just a relatively better price, which is important because our mission is the intelligence
06:09with everyone from day one. So we want to deliver a frontier intelligence model. At the same time,
06:15we hope everyone can get access to it. So the model should be strong, fast, stable, and affordable,
06:23accessible to everyone. This is our mission from day one. Okay, now just on the listing itself,
06:29so my understanding is, so you listed in Hong Kong. What about raising money or listing in mainland
06:36Chinese markets? If you could give us a sense of direction, rough direction, if that's in your
06:42horizon. I think it's all about like intelligence with everyone. So we want, currently you are a
06:49company. We think more and more people can get access to the company's model and can know about the
06:55company. It can probably become our shareholders to achieve AGI all together with us. Okay, how close
07:01do you think we are to AGI? Oh, that's a great question. It's based on your understanding on AGI. So
07:07based on our understanding, if the AI can self-generate 1% of the GDP, that probably became the AGI.
07:16I don't
07:16know what's the date, but I'm hoping for that it's coming soon. So we're probably not at 1% of
07:21GDP. Yeah,
07:22it can start it with like 1% of GDP and up to other higher percent. So the AI self
07:28-evolving
07:29and the self-planning execution and also doing evaluate for itself, it starts to happen. Yeah.
07:37Is it fair to say then that adoption is the number one priority then? I think intelligence is the number
07:45one priority. And then token consumption, token volume, as you said, adoption is important.
07:51And also growth profit margin will become also more and more important. What about chip constraints?
07:57Oh, I think all the frontier... Yes, tricky question. So all the frontier labs are facing the
08:03resources constraint, especially the chip constraint. So that's why we pay more attention on our technical
08:10leadership in model architecture, infrastructure, and optimization. So we invented, for example,
08:17we invented to minimize sparse attention, which drives up 10 times of the speed in inference,
08:22comparing to regular architecture, which will definitely accelerate our learning
08:29process. Right. And also, it will require less compute. Also, we design our model architectures,
08:38will regardless of chips. This is deliberate design principle from the beginning. So provide the
08:46frontier models with a competitive price that definitely drives the company's success.
08:52Earlier on the show, we were joined by Baidu a few minutes ago. You guys are among the new
09:00generation of pure AI labs and pure AI place. You're looking at the traditional, the bigger Chinese
09:06companies. Tencent, Alibaba, Baidu, what have you, ByteDance. They all have also very big budgets
09:13coming through. How do you look at competition now between the new generation of tech talent like
09:19yourself and the older generation? How do you compete with resources as much as we're looking at
09:28with the big tech talent? Actually, we collaborate a lot. Yeah, so I think... But they're also trying
09:35to get your market share. The key strategy is in our name, as I mentioned, Minimax. Minimaxing the cost,
09:42maximizing the intelligence. So you can see, based on our technical innovation from like pre-training,
09:48post-training, and like inference, we have a really strong competitiveness in both delivering
09:55frontier models and with the price leadership. So that will drive the more and more users that are using our
10:01models.
Comments