# Kimi’s open weights are closing the frontier gap

Markdown copy of https://realalpha.app/s/kimis-open-weights-are-closing-the-frontier-gap/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Alexander Wissner-Gross](https://realalpha.app/people/alexander-wissner-gross/) (Host) · Founder, Reified · Has backed and advised 40+ tech companies, directly and through 021T Capital and Reified; writes The Innermost Loop.
- Host: [Peter Diamandis](https://realalpha.app/people/peter-diamandis/) · Founder, XPRIZE · Co-founded Singularity University and founded Zero-G; co-wrote the books Abundance and Bold.
- Host: [Salim Ismail](https://realalpha.app/people/salim-ismail/) · Founder, Open ExO · Co-wrote Exponential Organizations; was Singularity University's founding executive director; sold Ångströ to Google.
- Host: [Dave Blundin](https://realalpha.app/people/dave-blundin/) · Founder & GP, Link Ventures · Has co-founded 23 companies; founding CEO of Vestmark and chairman of insurance marketplace EverQuote.
- Show: Moonshots
- Episode: [Anthropic’s $2T IPO, Elon Predicts Neocloud Death, and China’s Kimi Catches Up | #306](https://realalpha.app/v/WXELsIk4o0g/), Oct 28, 2026
- Moment: 54:32 to 1:02:59 of the episode (8 min), playing on https://realalpha.app/s/kimis-open-weights-are-closing-the-frontier-gap/; the episode from there: https://realalpha.app/v/WXELsIk4o0g/?t=3272

## Quote

> “China may have actually **passed the American public frontier**.”
> — [Alexander Wissner-Gross](https://realalpha.app/people/alexander-wissner-gross/) at [58:36](https://realalpha.app/v/WXELsIk4o0g/?t=3516) · Moonshots

## Takeaways

1. Kimi is reported to be only weeks behind the U.S. frontier.
2. Open weights can put competitive pressure on closed labs.
3. Distillation helps, but does not replace independent innovation.

## About this segment

The hosts examine Kimi’s benchmark results, low cost, and planned open-weight release. They argue that fast-following open models pressure frontier labs and give other countries a path to build capable AI.

From “Anthropic’s $2T IPO, Elon Predicts Neocloud Death, and China’s Kimi Catches Up | #306”, Oct 28, 2026. [Watch the full episode](https://realalpha.app/v/WXELsIk4o0g/)

## Transcript

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

[54:45](https://realalpha.app/v/WXELsIk4o0g/?t=3285) Over the last few days, we've seen the release of two major frontier models, Kami K 3.1 Pro and GPT 6.1 Terra, as well as about an hour ago, Google's Genie 4. that came out. Today, we're going to talk about Kimmy and Tara. We'll talk about Google Gemini Genie 4 next time. So first, let's cover Kimmy K3.1 Pro from the Chinese lab Moonshot AI. No relation to this podcast, but it is a great name. So Kimmy K3.1 Pro is a giant 2.8 trillion parameters with 104 billion active at any time and 1 million token context memory On Artificial Analysis Intelligence Index, Kimi scored 55.

[55:30](https://realalpha.app/v/WXELsIk4o0g/?t=3330) That put it fourth in the world behind Claude Opus 5.5, Claude Sonnet, and Metamuse Wonder. And it's cheaper. $2 per million tokens in, $10 out, a third of the last Kimi model. So here's why it matters. You know, next week, Moonshot AI is planning to release the weights, meaning anyone can download and run it. Once again, the gap between the best open models and closed models is shrinking from 12 points to three.

[56:02](https://realalpha.app/v/WXELsIk4o0g/?t=3362) Friend of the pod, Ramez Nam, ran the numbers and says, Kimi is now about five weeks behind the U.S. frontier. That's the shortest open closed AI gap has ever been. Let's take a look at this chart over here that Ramez put forward. um you know let's go to you uh imad what's your thoughts on this uh are we going to see a parody very soon so you know at times i've been a bit poo-pooing of distillation this is obviously distillation of opus 5.5 like if you look at the capabilities that it kind of did and anthropic they're extending the kimmy 3 model which was an interesting model that we discussed before The Anthropic S1 actually had a very interesting statistic.

[56:45](https://realalpha.app/v/WXELsIk4o0g/?t=3405) Anthropic banned 7% of their users. Wow. Accounts, which is probably, again, these distillation attacks that have been done on there. So I think that you're reaching this level of kind of performance parity and you've got the new QN models and new series of models coming out. You're just going to see a bunching up here, except for maybe the European labs. They're finding it difficult to catch up, but definitely from everyone else.

[57:07](https://realalpha.app/v/WXELsIk4o0g/?t=3427) They're all getting the same capabilities. One of the very interesting ones on this is that it can write these 3d files for games directly without going through blender or unreal or something like that so you can tell it like make a 3d picture a 3d model of salimus mail and it will actually start building it directly writing the code because it's learned from the representation of an opus 5.5 make a salim render to it going to blender to the output and this goes to what alex was just talking about it can go straight to the code now on things like that because of the way it's being distilled so you're getting some very funny things from distillation here alex here's the performance charts here's the benchmarks um i found it interesting it's at the top of humanity's last exam uh humanity's last

[57:56](https://realalpha.app/v/WXELsIk4o0g/?t=3476) exam is cooked, long cooked. I do think, if anything, we're underselling just a bit what an accomplishment this is from Moonshot, no relation. It's not just catching up on a raw capability term, raw capability terms to the U.S. frontier, which right now, at least by AAII benchmark standards, is Opus 5.5. It's actually, if you look at it on a performance per cost per task, it appears, at least to my eye, it's actually past the optimal frontier surface of Opus 5.5. So distillation, I suspect, based on allegations, is probably a large portion of the story. But there's something in there that seems to have passed Opus 5.5 in terms of cost or performance per cost, which is extraordinary. In that maybe spiky slash subjective sense, China may have actually passed the American public frontier. And if to the extent that is the case

[58:55](https://realalpha.app/v/WXELsIk4o0g/?t=3535) or at minimum caught up to it, I think this is extremely helpful for keeping anti-competitive AI safety cartels in the U.S. at bay, preventing them from pacing each other to have these Chinese models out there soon with open weights that are forcing them to compete, making them dance. And I think this is ultimately a net positive for the U.S. and the world. Love that. Dave? Yeah, I mean, it's incredible. The world we're living in right now, there's an alternate world where, you know, in Dario's fantasy world, there's no Chinese competitive pressure.

[59:30](https://realalpha.app/v/WXELsIk4o0g/?t=3570) And he would have stopped at Fable 5.1. He didn't even want to put out Fable 5.1. But because of the competitive pressure, he just had to put it out the door. Now it gets memed, you know, all over the place and picked on by Elon. But the thing is that that was the first model that was capable of improving itself. Now you're up to Office 5.5 and you've got Kimi K3 Pro clearly capable of self-improvement and they're out and they're available.

[59:53](https://realalpha.app/v/WXELsIk4o0g/?t=3593) And so that changes the roadmap for every large corporation with. any pride whatsoever, every sovereign nation with any pride whatsoever. There's a roadmap now where the exact cookbook from Kimmy is open source. You can actually see how to do it. And if you can get a toehold with a model like that, then you can use it to improve itself. But you have to take it seriously. You can't just play around because these things are, you know, they're improving so quickly now that, you know, either get in the game or you don't get in the game. But if China hadn't done that, we would be completely behind closed doors.

[1:00:27](https://realalpha.app/v/WXELsIk4o0g/?t=3627) There'd be three or four model providers. They would be accelerating it, but not releasing. You know, they'd be using it internally. You know, you'd have to join the company to get access to it. The flood of MIT talent to Anthropic would be even higher because everyone wants to just get their hands on it. But the Chinese have changed the course of history. And I agree. Alex is right. I want to stay away from the narrative that they just copied. Sure, they distilled, for sure.

[1:00:51](https://realalpha.app/v/WXELsIk4o0g/?t=3651) And distilling helps a lot, but you have to innovate. You know, they don't have the source code from Anthropic. And distilling gives you a really good roadmap to build toward, but it doesn't give you the answer key. So they've really innovated independently of the distillation. And so the combination of those two things, yeah, they're getting closer and closer and closer in time, you know, just a few weeks behind the frontier.

[1:01:15](https://realalpha.app/v/WXELsIk4o0g/?t=3675) And, you know, the curve looks very, shockingly consistent. So there's no reason to believe that that next iteration won't keep going up. And I'd love to get Alex and Ahmad's thoughts on, look, the y-axis on that chart. You know, we're showing intelligence go to the moon. But people really struggle to translate that into what is that going to look like? You know, what use case is going to crack? You know, what does it mean at 60 versus 70?

[1:01:38](https://realalpha.app/v/WXELsIk4o0g/?t=3698) Because that's coming very, very soon. So you're asking, Dave, what does the post-singular state of the world look like? I would say- and dirty though you know what is 50 versus 60 versus 70 on this y-axis what does that feel like 50 feels like a decent employee you have to keep an eye on 60 is an incredibly competent one and 70 is a superstar i think that's the best way to think about this and so we're about to crack great is there any world in which the open models exceed the closed models and capacity capability of course. Sure. In a world where the Western frontier labs decide to keep their best capabilities purely internal for legal or anti-competitive reasons.

[1:02:27](https://realalpha.app/v/WXELsIk4o0g/?t=3747) which may or may not be the world that we're headed towards, sure. I suspect we're certain. I think kind of by definition, though, if the open models bypass the closed models, then the closed people will use the open models. So they can't really shoot past the internal best. I think when you get to the highly competent employee and then the genius employee, which is imminent,