Liquid AI adds sub-1GB reasoning model

Liquid AI announced LFM2.5-1.2B-Thinking, a 1.2-billion-parameter model that writes out its reasoning before it answers. It is the newest member of the LFM2.5 family, which we covered at launch. The model is available on Hugging Face, Liquid's LEAP platform and its Playground.

In Liquid's own benchmarks, the Thinking variant scores well above its Instruct sibling on several tasks. It rises from 63.2 to 87.96 on MATH-500, from 60.98 to 69.33 on Multi-IF and from 49.12 to 56.97 on BFCLv3. Liquid says the model matches or beats Qwen3-1.7B in thinking mode on most reasoning benchmarks with 40% fewer parameters, and that it needs fewer output tokens to do so. Qwen3-1.7B still leads on MMLU-Pro (56.68 vs 49.65) and AIME25 (36.27 vs 31.73). Liquid recommends the Instruct model for chat and creative writing.

The training write-up focuses on "doom looping," where a reasoning model repeats itself and never reaches an answer. During preference alignment, Liquid always used a looping sample as the rejected response whenever one appeared. During RL with verifiable rewards (RLVR), it added an n-gram repetition penalty. On the company's test prompts, the share of looping outputs fell from 15.74% after midtraining to 0.36% after RLVR. The RL pipeline is an internal fork of verl that uses GRPO-style optimization. Instead of training on all domains at once, Liquid trained separate checkpoints for reasoning, math and tool use.

Qualcomm, Ollama, FastFlowLM and Cactus Compute join AMD and Nexa AI as launch partners.