Liquid AI reveals 8B-A1B MoE trained on 38T | Hacker News

Liquid AI unveiled an 8B-A1B MoE model trained on 38T tokens, underscoring rapid gains in small language models.

The Hacker News discussion says smaller models are becoming much more capable, especially with reinforcement learning and fine-tuning.

Models like Qwen3.5:4B are seen as increasingly strong options for local AI use.

The trend matters for privacy, speed, and lower hardware requirements as capable AI shifts onto local devices.