RT @LiquidAI_: We’ve partnered with @dottxtai and @remilouf to make function calling viable on the edge. JSON doesn’t cut it. We need stru….

@remilouf It's super fun to use a BERT-sized model that can run on a CPU and power a smart home interface. Also, we ranted a bit about function calling templates with a comparison between JSON and Python. Hope you enjoy it! :)

LFM2 🤝 Outlines. Constrained generation is amazing to squeeze more performance from tiny models. We just released a blog post with @remilouf with a smart home control demo. We used LFM2-350M to save notes, close blinds, and handle displays.

Oh no, it's a non-reasoning model. So it's even more interesting, but for the opposite reasons. Sorry about this mistake. :(.

Alright, this is just frontier models. Not ranking the companies or products.

Very nice post-training lesson from the new Qwen3 model:. - A pure reasoning model is expected to perform better on pure reasoning benchmarks. Nothing surprising. - We previously saw that pure reasoning models also perform very well on non-reasoning tasks, like creative writing.

I'll rate your incorrect rankings

Listen, I don't make the rules. But this is the only correct ranking.

I really like Kimi K2's writing style, but it completely hallucinates errors in my article review haha. reshape(-1, 1) isn't mentioned a single time 👀

I recommend reading the blog post to learn more about healthcare AI.

Apollo is a banger, congrats guys! 👏.

So proud of the WaifuGAN guys for fulfilling their dreams after six years and a research department

RT @LiquidAI_: Today, we release LEAP, our new developer platform for building with on-device AI — and Apollo, a lightweight iOS applicatio….

Try LEAP

We cracked the code to develop apps powered by local LLMs. We created LEAP based on our experience making end-to-end apps for customers. It's model-agnostic, developer-friendly, and available for iOS and Android today. Let's build the future of edge AI 🚀

@axolotl_ai @winglian 💻 Notebook: 🤗 LFM2-1.2B-Pirate:

You can now fine-tune LFM2 with @axolotl_ai! 🥳. Thanks to @winglian for this! You can use his notebook to smoothly SFT an LFM2 model. I tried it and created my own LFM2-1.2B-Pirate based on this code.

RT @ngxson: Trying out LFM2 350M from @LiquidAI_ and was mind-blown 🤯. The responses were very coherent. Less hallucinations compared to mo….

Hahaha 238 tokens/s prefill.

None of this scaffolding will die at the edge. Routers will only get increasingly relevant because edge models are catching up. Keep building smart and local solutions!.