Why Open Source AI Lags Behind China | Ioannis Antonoglou artwork

Why Open Source AI Lags Behind China | Ioannis Antonoglou

MTS

October 9, 2026

Reflection AI CTO Ioannis Antonoglou discusses Beam, their new 501B parameter open-weight model trained on 10,500 GPUs with zero distillation. He breaks down their massive RL scaling run, agentic generalizability, and how Western labs can compete with China's frontier model velocity.

0:00

Ioannis Antonoglou

There is a correlation between the performance of a model and how big it is, right? It's easier to compensate for a weaker model by just making it bigger. But then what happens is that you actually pay this cost during inference, because you have to serve a larger, more expensive model, but you don't really get the extra capabilities boost that would actually justify the extra compute.

0:27

Sophia

Hi everyone, and welcome back to MTS. Today, we are joined by Ioannis Antonoglou, who is the co-founder and CTO of Reflection AI. And this week, Reflection introduced Beam, a new 501 billion parameter AI model built for coding, reasoning and AI agents with plans to release its weights publicly later this month. So Ioannis, welcome to MTS.

0:50

Ioannis Antonoglou

Thank you so much for having me and really nice to meet you, Sophia.

0:53

Sophia

Really great to meet you as well. I wanted to talk a little bit about Beam. I know you just introduced this. This is the first frontier open weight model. Tell me a little bit more about why did you guys actually set out to build Beam and what can it do that you think is particularly important that you want the public to know about?

1:11

Ioannis Antonoglou

So the mission of the company is to build a frontier open intelligence. We believe that intelligence is probably the most important technology that humanity will ever build or it is built up to this point. We just want to ensure that there's a variety of options and we want to ensure that this technology is democratized and it's not just in the hands of a few closed labs. So to this end, we set out to just build frontier open intelligence and Beam is the first milestone on our path to achieving our mission. So Beam is really strong for its model class, 501 billion parameter model, it has 23 billion active parameters, which is a mixture of experts architecture. So at each forward path, we have 23 billion parameters that actually are activated. That's what it means.

We've actually trained it in a way so that it's really good at agentic tasks and it's really efficient with its reasoning. So what we want to see with Beam is a wider option of model, especially for use cases that you find in enterprises, for developers and if you actually take a look at the technology and the models that are mostly used in those settings, there are actually models at that parameter count, 500 billion parameters, because this is the model that you can easily serve internally to actually run inference on it. And the other thing that's really important is that you squeeze the most out of the computer you have available. So this is what we mean when we say that this model is exceptionally good at being reasoning efficient. What that means is that it makes really good use of these tokens so it just achieves tasks with far fewer tokens than other comparable models. And because it's also compact, it's smaller in size, 500 billion per hundres only, that means that it's also flop efficient. So the amount of compute that you actually need to spend to solve a particular task is low for the capabilities of the model. And again, we kind of did that because we felt that, especially in compute-constrained environments, it is really important that you just make the most of your compute, and you have something that's both compact but also highly efficient.

3:24

Sophia

Yeah, that makes a lot of sense.

I'm also curious too because when you guys set out to build Beam, one of the big things was focusing on autonomous coding, focusing on reinforcement learning, and you realized, this is something that was mentioned, is that the Western open model ecosystem doesn't have strong enough base models, and so that was one of the big reasons for actually pre-training your own. I'm curious if you've learned anything about just why Chinese open model labs have been so successful and what it would take for us in the West to actually be able to compete with them.

4:00

Ioannis Antonoglou

Yeah, so that's a really good question. So it is true that when we first started Reflection, we really believed in reinforcement learning and how reinforcement learning can actually unlock the next set of capabilities. And reinforcement learning is something that we really invested heavily in Beam. So Beam, for Beam, we actually had the largest reinforcement learning run that has been documented in the Open. We used 10.5 thousand GB 300s for four weeks. And this is what allowed us to just build a model that's so efficient. Because with reinforcement learning, you can ensure that your model becomes really, develops the right agentic reasoning. And by really scaling reinforcement learning, we managed to just get this really strong coding and agentic capabilities. And the other thing to note here is that the Chinese labs have been at it for like the past three years now. Like DeepSec is DeepSec before, because there have been like three more versions before that. Kimi is Kimi K3. So GLM is GLM 5.2 and 5.3. So you actually see that for this model to actually really catch up with the frontier, it just like takes a few durations.

Read the other 17 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000794054625