Why Frontier AI Models Fail at Biology | Alexander Amini
MTS
October 14, 2026
Alexander Amini, Chief Science Officer of Liquid AI, discusses the launch of their 24B MoE model (LFM 2.5), local vs. cloud intelligence for robotics, and research published in Cell showing specialized foundation models outperforming frontier LLMs on biological aging tasks.
Alexander Amini
If you take all of the frontier models, OpenAI, ChachiBT, Claude, Gemini, all across the board, these frontier models actually really struggle with this task. They are not very good predictors over this biological language, which is different than natural language that we as humans speak.
Sophia
Hello everyone, and welcome back to MTS. Today, I am joined live with Alexander Amini, who is the co-founder and chief science officer of Liquid AI and an AI scientist at MIT. And earlier, they also launched a new on-device AI model today called the LFM 2.5. So Alexander, welcome to MTS.
Alexander Amini
Thank you so much, Sophia. Glad to be here.
Sophia
Glad to have you as well. I'm super excited to talk a little bit more about this launch, what you guys officially announced today. Can you tell me more about this model and what was launched?
Alexander Amini
Absolutely. So this is our LFM 2.5 series of models. This is an extension of our LFM 2 series, which are these extremely efficient Liquid Foundation models that are trained for all different types of modalities. So we have LFM 2.5s that are for text language models, but also VL models, video language models, or vision language models, as well as audio language models. And today's release was the continuation of that series. It was actually our last model in that series. It was the LFM 2.5 24b MOE model. So it has 24billion parameters total, with 2billion parameters active. So it's an extremely efficient. It runs at the speed of a 2billion parameter model, but it has a much higher learning capacity. It's a very intelligent model as well, especially on a lot of agentic traces.
Sophia
Yeah, no, I'm particularly excited about this because I'm a huge fan of running my AI agents locally and also just running models locally. But the pushback I personally get is, oh, a lot of these local models just aren't that effective. They can't do the same things that maybe larger cloud models can actually do. And so I wanted to just get your take on what can this new model do maybe on a laptop that smaller models that typically also could run on a laptop couldn't do.
Alexander Amini
Absolutely. So I think, yeah, this is an excellent point. And it's one that we often see a lot is that big models are amazing generalists, right? So in the cloud, we have incredible frontier capabilities across a very wide spectrum of different types of tasks. But it also means that they're very poor specialists in the cloud as well. So what we really optimize at Liquid and with a lot of our Liquid foundation models is much more specialized capabilities. And what we show is actually that these models, when trained to be specialists, can even outperform the big models in the cloud, because those are almost trained to do many, many things very well. But as a result, it almost means that they can do not one thing very, very well at the same time, right? So it's kind of a trade off between exploration across this broad landscape and also exploitation across a single type of task. So in particular, this new LFM series that we released, this new LFM model that we just released, this 24 billion parameter model, this is an extremely strong agentic model. So it can really, it is made for AI PCs actually. It can fit in about 16 gigabytes of RAM, which many laptops have, and they can also be very, very strong for tool calling applications. So imagine you are doing agentic types of tasks. This is the type of model that we really optimized for in this one. Yeah.
Sophia
No, I really like that you say that, especially when it comes to tools and when it comes to specialized purpose reasoning, rather than just having a large model that can do a lot of generalized tasks. I want to get into what you actually researched with the cell paper in a second, because these are some interesting findings that I want to make sure that the MTS audience gets to hear about and understand. But before we just jump into that, I wanted to just get more of your takes around local models and just better understand who are the types of people who are adopting them and why are they adopting them?
Alexander Amini
Incredible question. So we see actually a few different types of categories of why people want local models. So there's a few different things. So number one is that we already talked about this generalist versus specialist point. A lot of people really like to prototype on big cloud models because they can do many, many things very broadly. But oftentimes after they know how they want to use a model, we oftentimes see a lot of interest in going local to basically specialize and narrow down frontier models into even higher than frontier capabilities just by specializing them. And that's where local becomes very important. Number two is actually in the connectivity regime. So if you care a lot about having your model run in disconnected environments where you don't need to be tethered to the Internet and you want to own your own intelligence, that's also a big area of where we see local models really starting to emerge. The reasoning there is that if you think of the majority of devices that we have in the world, and we all have devices all around us, right? So we have phones, laptops, cars, all of these are very intelligent or could be very intelligent computing devices.
Read the other 10 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:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown 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:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000796312971