Everything I Learned Training Frontier Small Models - Maxime Labonne, Liquid AI (AI column)
I watched this short one. report Maxime Labonne, Head of Post-Training, Liquid AI, where Maxim talks about how Liquid AI designs and trains edge models from architecture to reinforcement learning. The main point of the report is that small models cannot be perceived as smaller versions of large models. They have a different product physics: they live on phones, laptops, machines, IoT and embedded devices; they rely on memory, latency and the quality of specific tasks, rather than the universal ability to work in chat mode.
The main ideas of the speech are as follows:
1Small model should be specialized A large frontier model can be “average good at everything”: code, text, reasoning, search, creativity, long context. The little model doesn't work that way. It has few parameters, little memory for facts, and less room for self-correction. Therefore, a good use case for a small model is to close the narrow repeatable workflow: extraction, structured outputs, function calling, tool use, private processing of user data on the device. Liquid AI in the post about LFM2.5-350M straight. writer the limits of its applicability.
2Settings can be spent in the wrong place. In the slides, Maxim shows that in some small models, the embedding layer can take up a huge proportion of parameters: 63Percentage Gemma 3 270M and 29% at Qwen3.5-0.8B. For a small model, this is very expensive: formally, there are many parameters, but the “useful capacity” of the model is much smaller. LFM2.5-350M for Liquid AI embedding layer 19The effective size is listed as 287M and this directly affects latency, memory and quality.
3️⃣ FLOPs are not enough - you need to profile on the target gland The report focuses on on-device profiling. Liquid AI does not design architecture in a vacuum, but rather looks at what operations are really fast on target devices like the Galaxy S24 Ultra and Ryzen HX. 370. In general, if you build an AI feature for a mobile device, browser, laptop or embedded, you can not choose a model only by leaderboard. You need to measure prefill, decode, memory footprint, quantization impact, cold start, battery and tool-calling stability on your runtime.
4More pre-training can work even on a small scale Intuitively, it seems that the 350M model will quickly be saturated with data. (There's even Chinchilla scale of law). But Liquid AI shows the flip side: LFM2.5The 350M received additional pre-training from 10T to 28T tokens and large-scale reinforcement learning. Officially. post Liquid AI says the model outperforms more than twice the larger models on a range of benchmarks in knowledge, following instructions and using tools.
5Doom looping - separate failure mode of small reasoning models A very interesting part is about doom looping: when a model gets stuck in repeating text instead of coming to an answer. This is especially frustrating for small reasoning models and agentic scenarios: the user waits for an answer, a tool call or an action, and the model begins to repeat pieces of reasoning. Liquid AI describes its way of dealing with this, which reduced the doom loop 15.74percentage 0.36%. As a result, it is necessary to evaluate not only accuracy, but also “behavioral” errors such as looping, empty answers, and so on.
6Small models are especially interesting in agentic systems The future doesn’t necessarily look like “one huge model answers everything.” More realistic architecture - router + set of specialized models + tools + evals + fallback on a larger model. A large model can plan, reason, and analyze complex cases. A small local model can execute plans quickly and cheaply.
#AI #LLM #SmallModels #EdgeAI #Architecture #Engineering #ML #Agents #DevEx #Management