Liquid AI released LFM2.5-350M: efficient 350M parameter model with scaled RL

On March 31, 2026, Liquid AI released LFM2.5-350M — a compact language model with 350 million parameters, additionally trained on an expanded data corpus and a large volume of reinforcement learning. As reported by MarkTechPost, the release challenges the established rule in generative AI about scaling laws — the notion that more parameters almost always means more "intelligence" in a model.

Why challenge scaling laws

For the past several years, the large language model industry has developed mainly along one scenario: to get a more capable model, you need to increase the number of parameters, the volume of training data, and computational resources. This logic — the so-called scaling laws — became the foundation of a race among leading laboratories to create increasingly large flagship models. The problem is that models with hundreds of billions of parameters are expensive to train and operate: they are difficult to run locally, on devices with limited resources, or where low response latency is important.

Liquid AI positions LFM2.5-350M not as an attempt to compete with flagship models on absolute size, but as a technical case study on "intelligence density" — that is, how many useful capabilities can be packed into a model of fixed, relatively small parameter size if investment goes not into scale but into quality and volume of subsequent training.

What changed in model training

The main technical levers by which Liquid AI achieves increased model capabilities without increasing the number of parameters are additional pretraining and large-scale reinforcement learning. The volume of pretraining data has grown nearly threefold: from 10 trillion to 28 trillion tokens. Such an increase in the training corpus with unchanged model size is a direct bet that a model of fixed capacity can be "retrained" more densely, extracting from a larger text volume more useful patterns than was previously possible with less data.

Key parameters of the release:

• Model name — LFM2.5-350M
• Size — 350 million parameters
• Volume of pretraining data — increased from 10 trillion to 28 trillion tokens
• Additional stage — large-scale reinforcement learning
• Positioning — a technical case study on "intelligence density" rather than an attempt to set a size record

The model name itself — LFM2.5-350M — immediately fixes its key characteristic: 350 million parameters are several orders of magnitude smaller than modern flagship models with tens and hundreds of billions of parameters. This contrast is precisely what makes the release technically interesting: Liquid AI is not trying to hide the model's modest size, but rather puts it front and center in the name as a starting point for the experiment — how many capabilities can be extracted from a fixed small number of parameters through more prolonged and higher-quality training.

The second lever is large-scale reinforcement learning applied on top of expanded pretraining. If pretraining primarily saturates a model with broad knowledge of language and the world, the RL phase is typically used to direct the model's behavior toward more useful, accurate, and manageable responses — this phase in recent years has become the key way to improve model quality without increasing its size.

Why compact models are back in focus

The release of LFM2.5-350M fits into a broader 2026 trend: alongside the race for the largest flagship models, there is a separate race for maximally capable compact models that can run on phones, laptops, embedded devices, and in environments with limited computational resources — without calling cloud APIs. For developers, this means lower inference costs, offline operation, and lower response latency, which is critical for real-time applications.

If Liquid AI's approach is confirmed in practice — that is, if a 350 million parameter model truly demonstrates capabilities previously available only to much larger models — this will become an argument that further progress in AI does not necessarily require endless growth in the number of parameters but can be achieved through higher-quality and more prolonged training of existing architectures.