May 29, 2026
Small model, big main-character energy
Liquid AI reveals 8B-A1B MoE trained on 38T
Tiny AI, huge chaos: fans cheer the speed while others ask where Ollama went
TLDR: Liquid AI unveiled a small new model designed to run quickly on regular laptops while handling long inputs and tasks more smoothly. Commenters split between excitement over its surprisingly strong results and a mini-backlash over missing Ollama support, with one joker dubbing it "Homeopathic AI."
Liquid AI just dropped its new small model, LFM2.5-8B-A1B, and the pitch is basically catnip for people who want fast AI on ordinary laptops. The company says it can handle long documents, make multiple tool calls, and run well even on consumer hardware. Translation for normal humans: this is supposed to be a lightweight brain you can use locally, without needing a giant server farm. The crowd immediately zoomed in on the biggest flex: Liquid says it trained this thing on a staggering 38 trillion pieces of text and tuned it to be faster, better with long inputs, and more friendly to languages beyond English.
But the real show was in the comments, where reactions swung from full-on hype to drive-by roasting. One user called it "fucking phenomenal" after testing a long transcript summary, basically giving the model a standing ovation for punching above its weight. Another was thrilled by the idea that "we can run this even on CPU," which in local-AI land is code for finally, my humble machine gets invited to the party. And then came the shade: one commenter instantly nicknamed it "Homeopathic AI," a brutal little joke about how tiny this thing is.
The spiciest mini-drama? Support for Ollama, a popular app for running AI models locally. One user demanded to know why there's no day-one support, wondering if this is a temporary hiccup or a sign people are ditching Ollama for rival tools. So yes, Liquid launched a model — but the comments turned it into a referendum on tiny AI, local power, and which software ecosystem is winning the room.
Key Points
- •Liquid AI released LFM2.5-8B-A1B, an 8B MoE edge model aimed at fast on-device tool calling and local deployment.
- •The model scales pretraining from 12T to 38T tokens, expands context length to 128K, and increases vocabulary from 65,536 to 128,000 tokens.
- •Liquid AI says the architecture retains MoE, GQA, and gated short convolution blocks while shifting the new model to reasoning-only output behavior.
- •The company describes tokenizer changes that improve efficiency for non-Latin scripts and reports compression gains across 16 languages, with notable gains in Hindi, Thai, Vietnamese, Indonesian, and Arabic.
- •To support longer contexts and reduce looping behavior, Liquid AI used additional midtraining, increased the RoPE base, and added preference optimization plus reinforcement-learning shaping rewards.