Beam targets strong reasoning with fewer tokens

Transcript

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

1. 8:03

   Well, I think that brings us to what you guys are doing. You guys are talking about Beam, your first model, which I think everyone in the AI community has been buzzing about. What is reflection doing? When are they shipping? And the moment's coming now. So you've got Beam, and it's around a 500 billion parameter model. On the open benchmarks, it seems state-of-the-art. But I'd be curious to hear from you all, and as we're talking, you're still kind of going through the final phases of validation and all this, but I'd be curious to hear from you all what you expect the reaction to this model to be.

2. 8:36

   So Beam is our first model, is our first open-weight model. We're very, very excited to release it. I think it's, you know, building a company is, I think there's a Reid Hoffman quote that it's building a plane while you're flying it, and the plane's been built. So I think we're very excited about that. It is indeed around 500B total parameter model, 23B active, and it advances the frontier of Western open models in a meaningful way, which we're excited about.

3. 9:11

   The things that we think we're proud of with this model is its particularly strong encoding and agentic benchmarks. We believe that the form factor and the strength in these capabilities make it a really strong workhorse model for enterprises, public sector. It has this kind of utility, but also efficiency around it that I think will be appreciated. The kind of other thing that we're really excited about is that it's a model that is leading in its reasoning efficiency relative to other open models.

4. 9:48

   And when we say reasoning efficiency, we mean how fast does it solve a task relative to the compute it has at inference time or relative to the tokens that it has at inference time. And, you know, Yannis and team did some incredible work to get that statement you wanted to share a bit of that. The form factor model is a 500 billion parameter model. It's 23 billion active. And if you actually think of what you want from a model that's an open model, and you want to ensure that it has wide adoption, especially from enterprises, you need something that is efficient, it is compact, and it actually really produces the most intelligence that you can get out of your flops.

5. 10:29

   And in order to do that, you need to have a reasoner. These are reasoning models. And the idea is to just use tokens to actually solve tasks. The way that you measure whether a model is more efficient or like equally efficient to another model in terms of like reasoning efficiency is by just like seeing how many tokens or like how much time, how much compute it actually takes to solve a task of like, you know, similar difficulty.

6. 10:54

   And what we've actually seen is that we have a model, we've built a model that can be three or four times more token efficient, reasoning efficient than some of the open frontier models like GLM 5.3, or like some of the best closed models like GPT-6 Luna. So the question then is like, how can you ensure that you can extract the most reasoning and the most intelligence out of like this small model?

7. 11:22

   And the way to do it is by really scaling reinforcement learning. So reinforcement learning is something that, you know, I've been always like really passionate about. It's like, you know, my, I've started like doing reinforcement learning back in the early days of reinforcement learning in 2012, 2013. And, you know, it's something that like both Misha and I always like been really true believers in reinforcement learning. We've been RLP'd and you can actually see now that by scaling reinforcement learning and what we've actually done at Reflection is that we scale reinforcement learning to the highest level of compute in terms of like what we know publicly available in comparison to other like open labs. We've actually like you know released the research and we can just like compare apples to apples. So really kind

8. 12:08

   of like a double down on the compute we put into reinforcement learning with the goal of ensuring that we extract the most juice, the most intelligence out of the model. And this is ultimately what like enterprises and, you know, developers and everyone cares about, like a compact model that