# Small models can catch frontier capabilities within a year

Markdown copy of https://realalpha.app/s/small-models-can-catch-frontier-capabilities-within-a-year/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Marc Andreessen](https://realalpha.app/people/marc-andreessen/) · Marc Andreessen is cofounder of a16z. · Co-founded Netscape and co-created Mosaic, then co-founded a16z, a venture firm backing AI, crypto and frontier tech.
- Host: [Erik Torenberg](https://realalpha.app/people/erik-torenberg/) · General partner, Andreessen Horowitz · Helped build Product Hunt; founded On Deck, Village Global and Turpentine.
- Show: a16z Podcast
- Episode: [Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI](https://realalpha.app/v/xRh2sVcNXQ8/), Jan 7, 2026
- Moment: 16:11 to 19:52 of the episode (4 min), playing on https://realalpha.app/s/small-models-can-catch-frontier-capabilities-within-a-year/; the episode from there: https://realalpha.app/v/xRh2sVcNXQ8/?t=971

## Quote

> “after six or 12 months, there's a **small model that's just as capable**.”
> — [Marc Andreessen](https://realalpha.app/people/marc-andreessen/) at [16:41](https://realalpha.app/v/xRh2sVcNXQ8/?t=1001) · a16z Podcast

## Takeaways

1. Small models can catch up to leading models within six to 12 months.
2. He cites Kimi as a smaller model matching GPT-5 reasoning benchmarks.
3. He expects smaller models to account for much of the model volume.

## About this segment

Marc Andreessen says small models can match leading capabilities after six to 12 months, then deploy far more cheaply. He expects a hierarchy from data-center-scale models down to ones embedded in devices.

From “Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI”, Jan 7, 2026. [Watch the full episode](https://realalpha.app/v/xRh2sVcNXQ8/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[16:16](https://realalpha.app/v/xRh2sVcNXQ8/?t=976) And then, by the way, that also gets to this other kind of question where there's different theories on it, which is basically big models versus small models. And so a lot of the data center build is oriented around hosting, training, and serving the big models, you know, for all the obvious reasons. But there's also the small model revolution is happening at the same time. And if you just kind of track, you know, you can get the various research firms have these charts you can get.

[16:42](https://realalpha.app/v/xRh2sVcNXQ8/?t=1002) But if you just kind of track the capability of the leading-edge models over time, what you find is after six or 12 months, there's a small model that's just as capable. And so, there's this kind of chase function that's happening, which is the capabilities of the big models are basically being shrunk down and provided at smaller size and then, therefore, smaller cost, you know, quite quickly. So, I'll just give you the most recent example that's just kind of hit over the last two weeks.

[17:05](https://realalpha.app/v/xRh2sVcNXQ8/?t=1025) And again, this is the thing that's just kind of shocking, is there's this Chinese company that has a, well, I forget the name of the company, but it's the company that produces the model called Kimi, just spelled K-I-M-I, which is one of the leading open source models out of China. And the new version of Kimi is a reasoning model that, at least according to the benchmark so far, is basically a replication of the reasoning capabilities of GPT-5, right?

[17:27](https://realalpha.app/v/xRh2sVcNXQ8/?t=1047) And the reasoning models of GPT-5 were a big advance over GPT-4, and of course, GPT-5 costs a tremendous amount of money to develop and to serve. And all of a sudden, here we are, whatever, six months later, and you have an open-source model called Kimi. And I think, I don't know if they've had, it's either shrunk down to be able to run on either, it's like one MacBook or two MacBooks, right? And so, you can all of a sudden, if you have like an application, if you're a business and you want to have a reasoning model that's GPT-5 capable, but, you know, you're whatever, you're not going to pay the whatever GPT-5 cost or you're not going to want to have it be hosted and you want to run it locally, you know, you can do that.

[18:00](https://realalpha.app/v/xRh2sVcNXQ8/?t=1080) And, again, that's just like another, just like another, you know, it's another breakthrough. Like, it's another Tuesday, another huge advance. It's like, oh, my God. And then, of course, it's like, all right, well, what is OpenAI going to do? Well, obviously, they're going to go to GPT-6, right? And, you know, right. And so there's this kind of laddering that's happening where the entire industry is moving forward.

[18:18](https://realalpha.app/v/xRh2sVcNXQ8/?t=1098) The big models are getting more capable. The small models are kind of chasing them. and then the small models provide, you know, completely different way to deploy, you know, at very low price points. And so, yeah, and, you know, we'll see what happens. I mean, there are some very smart people in the industry who think that ultimately everything only runs in the big models because obviously the big models are always going to be the smartest, and so therefore you're always, you know, you're always going to want the most intelligent thing because why would you ever want something that's not the most intelligent thing for any application?

[18:45](https://realalpha.app/v/xRh2sVcNXQ8/?t=1125) You know, the counterargument is just there's a huge number of tasks that take place in the economy and in the world that don't require Einstein, you know, where, you know, 120 IQ person is great. You don't need a, you know, 160 IQ, you know, PhD in, you know, string theory. You just like have somebody who's competent and capable and it's great. And so, you know, I, you know, and we've talked about this before.

[19:04](https://realalpha.app/v/xRh2sVcNXQ8/?t=1144) I tend to think the AI industry is going to be structured a lot like the computer industry ended up getting structured, which is you're going to have a small handful of basically the equivalent of supercomputers, which are these like giant, you know, kind of what we call god models that are, you know, running in these giant data centers. And then, you know, I'm not like convinced on this, but my kind of working assumption is what happens is then you have this cascade down of smaller models, ultimately all the way to very small models that run in embedded systems, right, run on individual chips inside every, you know, physical item in the world.

[19:33](https://realalpha.app/v/xRh2sVcNXQ8/?t=1173) And that, you know, the smartest models will always be at the top, but the volume of models will actually be the smaller models that proliferate out. And right, that's what happened with microchips. It's what happened to computers, which became microchips. And then it's what happened with operating systems and with a lot of everything else that we built in software. So, you know, I tend to think that's what will happen.