Liquid AI releases LFM2.5-350M: a 350M-parameter model trained with large-scale RL

On March 31, 2026, Liquid AI released LFM2.5-350M, a compact language model with 350 million parameters, further trained on an expanded data corpus and a large volume of reinforcement learning (RL). As MarkTechPost reported, the release challenges a well-established rule in generative AI, the scaling laws: the idea that more parameters almost always mean more "intelligence" in a model.

Why challenge scaling laws

For the past several years, the large language model industry has mostly followed the same playbook: to get a more capable model, increase the number of parameters, the amount of training data and the compute. This logic, known as scaling laws, underpinned the race among leading labs to build ever larger flagship models. The problem is that models with hundreds of billions of parameters are expensive to train and to run: they are hard to deploy locally, on resource-constrained devices, or where low response latency matters.

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

What changed in the model's training

The main technical levers Liquid AI uses to improve the model without increasing its parameter count are additional pretraining and large-scale reinforcement learning. The pretraining data grew almost threefold, from 10 trillion to 28 trillion tokens. Expanding the training corpus while keeping the model size unchanged is a direct bet that a model of fixed capacity can be trained more densely, extracting more useful patterns from a larger body of text than was possible before with less data.

Key release details:

• Model name: LFM2.5-350M
• Size: 350 million parameters
• Pretraining data: grew from 10 trillion to 28 trillion tokens
• Additional stage: large-scale reinforcement learning
• Positioning: a technical case study in "intelligence density", not an attempt to set a record for model size

The name itself, LFM2.5-350M, states the model's defining trait up front: 350 million parameters is several orders of magnitude smaller than today's flagship models, which have tens or hundreds of billions. That contrast is what makes the release technically interesting.

Liquid AI does not try to hide the modest size; it puts it in the name as the starting point of the experiment: how much capability can be extracted from a fixed, small number of parameters through longer and better training. The second lever is large-scale reinforcement learning, applied on top of the expanded pretraining. Pretraining mainly fills the model with broad knowledge of language and the world, while the RL stage is typically used to steer the model toward more useful, accurate and controllable answers.

In recent years, this stage has become a key way to improve model quality without increasing size.

Why compact models are back in focus

The LFM2.5-350M release fits a broader 2026 trend: alongside the race for the largest flagship models, there is a separate race for the most capable compact models, which can run on phones, laptops, embedded devices and in compute-constrained environments, without calling a cloud API. 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 holds up in practice, meaning that a 350-million-parameter model really does show capabilities previously available only to far larger models, it would support the argument that further progress in AI does not necessarily require endless growth in parameter counts, and can come from better and longer training of existing architectures.