Small models need task-specific training because their memory limits reduce their knowledge capacity.
2
Liquid AI uses on-policy preference alignment and reinforcement learning to reduce doom loops in small reasoning models.
3
Small models become more useful when they can call tools such as web search instead of relying on stored knowledge alone.
Summary
Maxime Labonne explains why small edge models need their own architecture and training approach. They are limited by device memory, need low latency, and usually work best on narrow tasks rather than as general-purpose chatbots. Liquid AI reduces wasted capacity in small models by using a smaller embedding layer and gated short convolutions, which Labonne says improve speed and memory use on target devices. The LFM 2.5 recipe combines continued pre-training, supervised fine-tuning, on-policy preference alignment, and reinforcement learning. Labonne focuses on doom loops, where a reasoning model repeats text indefinitely. Preference data can reject looping rollouts, while verifiable rewards and repetition penalties make loops almost disappear after reinforcement learning. He argues that web search, Python, and other tools can compensate for limited knowledge and context, making small models especially suitable for private, offline, low-latency, and agentic workloads.
Small models trade general knowledge for memory use, focus, and speed
Labonne describes small models as memory-bound because they must run on devices such as phones and cars. Their limited memory gives them less knowledge capacity than larger models, so they usually work better on a narrow task than as general-purpose chatbots. Summarization is one example. They are also latency-sensitive, which means inference needs high throughput. These constraints interact, and Labonne says small models are not simply scaled-down versions of larger models. Their training and design have their own problems.
Large embedding layers waste capacity in very small models
Labonne compares the smallest Gemma models and points out that their embedding layers take up a large share of the parameter budget. In Gemma 3 270M, the embedding layer is 63% of the total parameters. In Gemma 2.5 0.8B, it is 29%. He calls the remaining parameters the effective parameters because they do the work related to reasoning and knowledge capacity. The large embeddings come from distillation setups with teacher models that have huge vocabularies. A smaller embedding layer could leave more of the same memory footprint for useful model computation.
LFM2 uses hardware profiling to choose fast short convolutions
Liquid AI designed LFM2 with a hybrid architecture that combines short convolutions and group query attention. Labonne says the team profiled models on the hardware where they would run instead of relying only on theoretical comparisons. This led to a gated short convolution block. In his comparisons, short convolutions have a lower cost than sliding window attention, gated Delta Net, gated linear attention, and group query attention. Profiles on an AMD Ryzen Max Plus 395 and a Samsung Galaxy S25 Ultra showed that LFM2 was faster and used less memory. He also says the architecture had strong GPU throughput at high concurrency.
More pre-training tokens can still help a 350M model
The LFM2.5 recipe includes pre-training and mid-training on 28 trillion tokens, followed by supervised fine-tuning, preference alignment, and reinforcement learning. Labonne notes that this amount may look unusual under Chinchilla scaling laws for a 350M model. He cites a recent paper by Roberts et al. on test-time scaling laws and says the models may benefit from even more pre-training. The LFM2.5 350M model improved over LFM2 350M across knowledge, instruction following, data extraction, and tool-use benchmarks. Liquid AI chose to focus the model on data extraction and tool use instead of making it average at every capability.
Narrow supervised fine-tuning gives small models a better starting point
Labonne recommends making supervised fine-tuning narrow and focused on the intended use case. Calling a particular function is an example of a good small-model task. Preference alignment has a wider effect in his description, improving the model beyond individual benchmarks and making its overall responses better. Reinforcement learning is highly efficient even at small scale, but it needs many environments and tasks to generalize. Small models are especially sensitive to cold-start SFT data. If a reinforcement learning task fails to train, adding similar examples to the supervised fine-tuning mixture may help.
Doom loops appear when small reasoning models face tasks beyond their ability
A doom loop happens when a model repeats a sequence of words and never stops. Labonne says the problem is especially common when three conditions meet: the model is small, it is a reasoning model, and the task is complex. A tiny reasoning model solving very difficult math problems is his example. He presents doom looping as a distinct training challenge for small models. In one LFM2.5 1.2B thinking example, the doom loop ratio across many benchmarks was about 15% to 16% after pre-training. Supervised fine-tuning barely changed it because the SFT data did not contain doom-loop examples.
On-policy preference data and reinforcement learning can remove doom loops
Liquid AI generates about 1 million prompts, then samples five rollouts from the policy model with temperature sampling. It also generates one rollout at temperature zero, which Labonne expects will often doom loop. An LLM jury scores the rollouts, and the best and worst responses become the chosen and rejected answers for preference alignment. A looping answer should therefore be rejected during training. The second method uses reinforcement learning with verifiable rewards and a small n-gram repetition penalty. For a math question, a rollout without a final answer receives no positive reward. Labonne says preference alignment reduces the loop rate substantially and reinforcement learning makes it almost nonexistent in the example.
Tools can compensate for limited knowledge and context
Because small models have low knowledge capacity, Labonne suggests giving them tools such as web search. A tiny model that can search for answers may perform better on knowledge questions than one that relies only on its parameters. He also says small models can be good at agentic tasks when their reasoning lets them use tools reliably. Their weaker long-context ability can be addressed with a recursive language model environment that uses Python to take shortcuts. His broader point is that many small-model limitations can be addressed by changing the surrounding system rather than treating the model like a large one.
Offline use, latency, and privacy are practical reasons to choose small models
In the question-and-answer section, Labonne gives three situations where small local models make sense. An in-car system may not have a reliable internet connection, so it needs to work offline. Latency-sensitive workloads benefit from a small model running locally. Privacy is another reason, especially in regulated areas such as finance and health care. When asked whether doom-loop improvements from reinforcement learning can be distilled from a larger model, he says the answer needs experiments. His expectation is that distillation alone would not fully solve the issue and that several training batches might still be needed.
"If you have a tiny model, but it's able to Google everything that you throw at it in terms of knowledge questions, you're going to have much much better performance than if you just rely on the base models."15:52
Who should watch
You are fine-tuning a small model for structured extraction, function calling, or another narrow task and need a training recipe.
Your reasoning model repeats itself on hard tasks, and you need practical ways to reduce doom loops.
You are choosing between a local model and a larger hosted model for an offline, low-latency, or privacy-sensitive application.
