AI
Small Models Face Unique Constraints
Liquid AI builds edge models from 350 million to 24 billion parameters. They released a 450M VLM yesterday and a new 350M text model last week. Small models are memory bound because hardware on phones and cars limits size. They have low knowledge capacity compared to big models. They are task-specific, not general-purpose chatbots. They are latency sensitive and need fast throughput. Small models are not scaled-down versions of bigger models. They have their own challenges.
Architecture: Embedding Layers and Short Convolutions
Gemma 3 270M is 63% embedding layer. Gemma 2.5 0.8B is 29% embedding layer. Those parameters do not contribute to reasoning. Liquid’s LFM 2 architecture uses a gated short convolution block. Short convs are faster than sliding window attention, gated Delta Net, gated linear attention, and group query attention. On an AMD Ryzen Max Plus 395 and Samsung Galaxy S25 Ultra, short convs deliver higher throughput and lower memory use. The embedding layer is only 10% of total parameters, leaving more room for effective reasoning.
Training a 350M Model on 28 Trillion Tokens
LFM 2.5 350M was pre-trained on 28 trillion tokens. That violates Chinchilla scaling laws, but performance still grows with more tokens. A new paper by Roberts et al. confirms that small models benefit from additional pre-training. Post-training includes supervised fine-tuning, preference alignment, and reinforcement learning. The model excels at data extraction and tool use. Benchmarks: GPQA Diamond, IF Bench, Case Report Bench, BFCL, and Dow 2 Bench. It is not average on everything. It targets specific capabilities.
Doom Looping and How to Fix It
Small reasoning models on hard math tasks repeat words endlessly. This is doom looping. LFM 2.5 1.2B thinking had a 15% doom loop ratio after pre-training. SFT barely moved it. Preference alignment (DPO) with on-policy data generation using temperature sampling and an LLM jury reduced it significantly. Reinforcement learning with verifiable rewards and an n-gram repetition penalty almost eliminated it. Qwen 3.5 0.8B in reasoning mode has over 50% doom loops. Liquid treats edge models as their own thing, not scaled-down big models.
Agentic Reinforcement Learning and Next Steps
Small models are memory bound and hallucinate. Giving them web search tools solves knowledge gaps. They are good at agentic tasks because they need reasoning, not knowledge. Recursive language model environments with Python handle long context issues. Liquid is working on LFM 3 with many experiments and ideas. The message: small model problems are fixable with creativity and a different mindset from big model thinking.
Q&A
How do you decide between small and big models in your workflow? Small models are used when there is no internet connection, for latency-sensitive workloads, or in regulated environments like finance and healthcare. ▶ Watch (18:00)
Can RL-based doom loop reduction on a big model be distilled into a smaller model? Probably not. Distillation is too close to SFT and would not completely solve doom looping. Several batches of RL would still be needed. ▶ Watch (19:19)
Notable Quotes
“Small models are not just scaled down versions of bigger models. They also have their unique challenges.” Maxime Labonne · ▶ Watch (2:07)
“If you have a tiny reasoning model on super difficult math task, this is the perfect recipe to have a lot of doom loops.” Maxime Labonne · ▶ Watch (11:18)
“Most of the issues that you find with small language models can actually be fixed in different ways. It just requires more creativity.” Maxime Labonne · ▶ Watch (16:54)
Key Takeaways
- Small models are memory bound, task-specific, and latency sensitive, not scaled-down big models.
- Short convolutions in LFM 2 architecture outperform sliding window attention on edge hardware.
- Doom looping in small reasoning models can be reduced from 15% to near zero with preference alignment and RL.