Everything I Learned Training Frontier Small Models — Maxime Labonne, Liquid AI

Maxime Labonne, Head of Pre-training at Liquid AI, presents a technical deep-dive into the unique challenges and solutions involved in post-training frontier small language models, drawing on lessons from developing the LFM2.5 model family (350M–1.2B parameters) designed for on-device deployment.

Labonne begins by distinguishing small models from large ones along three axes: they are memory-bound (constrained to under 1 GB on phones, cars, and edge hardware), task-specific rather than general-purpose, and extremely latency-sensitive. He argues that small models are not simply scaled-down large models — they require a fundamentally different engineering mindset.

On architecture, Labonne critiques the inefficiency of large embedding layers in models like Gemma 3 270M (63% of parameters are embeddings) and Gemma 2.5 0.8B (29%), attributing this to distillation from large-vocabulary teacher models. In contrast, Liquid's LFM2 architecture uses on-device profiling on real hardware — an AMD Ryzen Max Plus 395 and a Samsung Galaxy S25 Ultra — to identify optimal operators. The result is a gated short convolution block that achieves dramatically lower latency and memory usage compared to sliding window attention, gated Delta Net, and GQA alternatives, while maintaining strong GPU throughput at high concurrency.

The LFM2.5 training recipe involves 28 trillion pre-training tokens for a 350M parameter model — far beyond Chinchilla-optimal, but Labonne cites recent work by Roberts et al. on test-time scaling laws showing performance continues to improve with more tokens even at the smallest scales. Post-training stages include supervised fine-tuning (SFT), on-policy length-normalized Direct Preference Optimization (DPO), and reinforcement learning with verifiable rewards (RLVR).

A central technical contribution is the identification and mitigation of 'doom loops' — infinite token repetition sequences that disproportionately affect small reasoning models on hard tasks. Labonne measured a 15–16% doom loop rate in LFM2.5 1.2B Thinking after pre-training, compared to over 50% in Qwen 3.5 0.8B (which he characterizes as merely a scaled-down large model). Two solutions are detailed: (1) on-policy preference alignment using a pipeline that generates five diverse temperature-sampled rollouts plus one greedy rollout, submits all to an LLM jury, and trains the model to prefer non-looping responses via DPO; and (2) RLVR with n-gram repetition penalties and temperature sampling, which nearly eliminates doom loops. After combining both stages, the doom loop ratio drops to near zero.

For SFT, Labonne emphasizes the importance of cold-start data that closely mirrors the RL task distribution; without it, RL training stalls and the model struggles. He recommends targeting narrow capabilities (e.g., structured data extraction, multi-turn tool use) rather than chasing average benchmark performance across all domains.

Looking ahead, Labonne argues that small models are underexplored as agentic systems. Their low knowledge capacity leads to hallucination, but web search and tool access effectively compensate. He contends that small models excel at agentic reasoning — they just need reliable tool-use capabilities, not encyclopedic knowledge. Recursive environments with Python access can also substitute for weak long-context performance. Liquid is actively working on LFM3 to push these ideas further.