Video yükleniyor...
Video Yüklenemedi
This is a 1.7B param model btw, using the browser, making html files, and managing itself in Pi. This model hits 4o levels on most benchmarks lol Great for workflow automations.
121,160 görüntüleme • 5 gün önce •via X (Twitter)
32 Yorum

Yeah this lfm model is amazing - works on my android with 3gb ram at 20t/s

nice - I hadn't tried the new LFMs with more agentic capabilities. going to look into training hybrid edge/cloud semantic routing for it in our next ARC checkpoint

4o levels? So great at causing psychosis but not much else?

pretty epic. Also the 2B with vision punches way above its weight

You can even use it on a Raspberry Pi: It’s the stock one, without webGPU kernels.

I’ve been experimenting with this one too. For laptops without great chips, it’s pretty amazing what it can do!

heard a bunch of terrible stuff about lfm3, but glad to see someone doing some good stuff with this will def look into lfm3 automation into a new agent focused browser im working on

tried it, trashed it

a 1.7B model writing its own html files would’ve sounded made up last year

What's the model?

No local model matches 4o in anything.

Let's add this on Singularity Network.

This is the size I care about for office automations. My 14B on a 3080 runs around 18 tok/s with some layers on CPU, a 1.7B that can drive a browser leaves the whole card for context. How's it hold up on longer multi-step stuff?

Yup, my poor man’s local agent 🥹 works like a charm👌

It's pretty good but it's quality of work could use work

Which model is that ?

Awesome!

Lfm 3 my guess i have tried it too

It is certainly one of the greatest models.
It's Mac browser I assume? My iPhone browser refuses to run even 325m param model - site. Runs great on Mac though. Feedback is most welcome!

fiquei curioso, essa interface bem feita, é de qual app? é nativa?

Gonna go check it out. AI for exploration is awesome.

this is great. any writeup or link i can read ? I am trying to run it in collaboration with picoclaw for example but it loses context also for simple action requests and conversations

Demo videos need bass-heavy instrumental dance soundtracks for me to take them seriously.

how do you recommend using a model like this?

How do you build custom silicon for Liquid AI’s LFM3-1.7B? The tri-hybrid architecture (Conv + Sliding Window + Global Attention) requires a completely different ASIC design than pure transformers. Here’s how to architect an edge AI chip for it.

Pretty insane this is even possible. How’s tool calling capability?

Liquid cooked with this one

on what inference engine do you run it ?

The browser part is what got me. I run a small model on my own machine for the boring jobs, and the moment it stopped needing the cloud my whole workflow got calmer. A 1.7B model that makes its own HTML files and manages itself is the first time the small end of the scale has felt like the useful end to me.

Can it reliably run locally on a Pi?

And it’s amazing at vision, I was completely stunned
