Liquid AI ships 38T-token MoE

Liquid is pushing the argument that small, sparse models plus aggressive systems work can beat brute-force scaling for real on-device agents. The interesting part here is less the raw parameter count than the combination of longer context, multilingual tokenizer expansion, and reasoning-oriented tuning.

• –38T-token pretraining is a major jump over the prior 12T run, so this is a substantive model rebuild, not a minor refresh
• –128K context makes it more viable for long-document work, multi-step tool use, and agentic workflows on local machines
• –Liquid is optimizing for practical deployment, with day-one support for llama.cpp, MLX, vLLM, and SGLang
• –The model is explicitly reasoning-only, which should help structured problem solving but also raises the usual edge-model tradeoff around knowledge breadth and hallucinations
• –If the benchmark claims hold up outside Liquid’s own eval stack, this is another sign that on-device AI is shifting from “can it fit?” to “can it reliably act?”