YHacker Newsscreenshots

Screenshot of liquid.ai: Introducing LFM3: Every Modality for Specialized Intelligence

First screen, captured 2 days ago. Hover to scroll the page, or open the video.

Introducing LFM3: Every Modality for Specialized Intelligence

More from liquid.ai →

Liquid AI blog post announcing LFM3, a family of open-weight ~2B-parameter multimodal foundation models with a new tri-hybrid architecture for on-device use.

What HN thinks

Tiny models, big hopes, and one cow that apparently says 'oink'.

Small local models for triage
Users already run earlier LFM versions for annotating data, home automation and classification, and see LFM3 as a cheap local upgrade, especially with fine-tuning.
Fine-tune for specific tasks
Commenters and a Liquid team member stress that these models shine when fine-tuned on modest labeled data for narrow jobs like OCR, action items or tagging.
Skepticism on rough edges
Some report odd reasoning traces, quantization-related wrong answers, MLX loading failures, and doubt such small models are meant for coding.

Consensus: Most agree LFM3 is impressive for its size and best suited to specialized, local, fine-tuned workloads rather than general frontier tasks.

Recommended in the comments

GLiNER (library): Fine-tunes in minutes and runs well on CPU for tagging tasks.
mlx-vlm ↗ (library): Suggested as the way to run the LFM3 Audio and Omni models on Macs.
Home Assistant (tool): Used with small local LLMs as a voice/home assistant backend.

Summarised by Claude from the comment thread 2 days ago.

Points: 146
Comments: 55
Best rank: #10
Posted: 2026-10-27 by Philpax
Category: Ai ml

liquid-ai lfm3 small-language-models multimodal on-device-ai open-weights edge-ai

All captures:

Full page

Full-page screenshot of liquid.ai

Captured 2 days ago. Click for the whole screenshot. The page is longer than 25,000px, so the screenshot stops there: read the rest on liquid.ai ↗

Looks similar

Screenshot of liquid.ai: LFM2-24B-A2B: Scaling Up the LFM2 Architecture

LFM2-24B-A2B: Scaling Up the LFM2 Architecture (liquid.ai)

52 points by nateb2022 6 months ago | 12 comments

Screenshot of liquid.ai: Reducing Doom Loops with Final Token Preference Optimization

Reducing Doom Loops with Final Token Preference Optimization (liquid.ai)

24 points by dataminer 3 months ago | 5 comments

Screenshot of liquid.ai: LFM2.5-24B-A2B: Built for fast on-device agents

LFM2.5-24B-A2B: Built for fast on-device agents (liquid.ai)

271 points by gmays 15 days ago | 72 comments

Liquid AI blog post announcing LFM2.5-24B-A2B, a sparse language model with about 2B active parameters designed for fast on-device agents.

liquid-ai lfm2-5 mixture-of-experts on-device-ai

Screenshot of liquid.ai: Liquid AI reveals 8B-A1B MoE trained on 38T

Liquid AI reveals 8B-A1B MoE trained on 38T (liquid.ai)

227 points by simjnd 5 months ago | 86 comments

Screenshot of dactyl.dev: How Dactyl Works

How Dactyl Works (dactyl.dev)

60 points by anorak27 2 months ago | 12 comments

Screenshot of jellyfin.org: Jellyfin LLM/"AI" Development Policy

Jellyfin LLM/"AI" Development Policy (jellyfin.org)

170 points by mmoogle 9 months ago | 90 comments

Related stories

Screenshot of modular.com: Mojo – a new programming language for AI developers

Mojo – a new programming language for AI developers (modular.com)

482 points by lairv 3 years ago | 185 comments

Modular's product page introducing Mojo, a new programming language for AI developers that combines Python usability with high performance.

mojo modular programming-languages python

Screenshot of stability.ai: StableLM: A new open-source language model

StableLM: A new open-source language model (stability.ai)

1412 points by davidbarker 3 years ago | 474 comments

Stability AI blog post announcing the release of StableLM, the first models in its open-source language model suite.

stablelm stability-ai large-language-models open-source-models

Screenshot of blog.google: EmbeddingGemma 2: An open, lightweight multimodal embedding model

EmbeddingGemma 2: An open, lightweight multimodal embedding model (blog.google)

395 points by ilreb 23 days ago | 42 comments

Google blog post announcing EmbeddingGemma 2, an open lightweight on-device model that embeds text, images, audio, and video into a unified space.

embeddinggemma gemma embeddings multimodal

Screenshot of github.com: FastVLM: Efficient vision encoding for vision language models

FastVLM: Efficient vision encoding for vision language models (github.com)

335 points by nhod 1 year ago | 68 comments

Apple's GitHub repository with the official implementation of FastVLM, an efficient vision encoding method for vision language models presented at CVPR 2025.

fastvlm vision-language-models apple cvpr-2025

Screenshot of thinkingmachines.ai: Inkling: Our Open-Weights Model

Inkling: Our Open-Weights Model (thinkingmachines.ai)

1146 points by vimarsh6739 3 months ago | 278 comments

Thinking Machines blog post announcing Inkling, the company's open-weights AI model, dated July 15, 2026.

thinking-machines inkling open-weights llm

Screenshot of sergey.fyi: AI agents: Less capability, more reliability, please

AI agents: Less capability, more reliability, please (sergey.fyi)

393 points by serjester 1 year ago | 230 comments

Blog post by Sergey Filimonov arguing AI agents should prioritize reliability over expanded capability, critiquing flight-booking demos.

ai-agents reliability llm product-design