--- title: Liquid AI announces LFM3, a new generation of Liquid... description: Liquid AI announced LFM3, the next generation of its Liquid Foundation Models. The company says the models use a new tri-hybrid architecture and new multimodal... canonical: https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Liquid AI announces LFM3, a new generation of Liquid Foundation Models | daily.dev og:description: Liquid AI announced LFM3, the next generation of its Liquid Foundation Models. The company says the models use a new tri-hybrid architecture and new multimodal... og:url: https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh og:image: https://api.daily.dev/og/posts/xBp2peUEH.png og:image:alt: Liquid AI announces LFM3, a new generation of Liquid Foundation Models og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Liquid AI announces LFM3, a new generation of Liquid Foundation Models **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Liquid AI announced LFM3, the next generation of its Liquid Foundation Models. The company says the models use a new tri-hybrid architecture and new multimodal encoders. ## Content Liquid AI has released LFM3, a new generation of its Liquid Foundation Models. It is a family of open-weight models around 2B parameters, covering text, vision, audio, omni and decision tasks on one shared hybrid backbone. Weights are on Hugging Face. The models are also available in the Liquid AI Playground and API. ## Architecture LFM3 uses what Liquid calls a tri-hybrid design. In the 1.7B text model, 13 of 24 layers are short convolutions, 5 use 256-token sliding-window attention, and 6 attend globally. The models were pretrained on 22T tokens. The payoff is efficiency. LFM3-1.7B beats the previous LFM2.5-2.6B with about 35% fewer parameters. It fits in roughly 1 GB at 4-bit precision and has a 256K-token context window. Liquid says it decodes 1.8 to 2.8 times faster than competing models on-device. ## Text model results LFM3-1.7B is trained to run in agent harnesses such as Hermes Agent and Pi. All figures below come from Liquid's own evaluations. - WebArena-Verified: 38.4, against 28.9 for Qwen3.5-9B and 3.02 for LFM2.5-2.6B. - AutomationBench: 41.29, against 15.10 for Qwen3.5-9B. - Long tasks in browsers and business apps: ahead of Qwen3.5-9B, a model more than five times larger. ## Vision model LFM3-VL-2B raises the average vision benchmark score from 57.6 (LFM2.5-VL-3B) to 71.4. Liquid says it beats same-size rivals by at least 7.4 points across 62 benchmarks, and it keeps pace with Gemma 4 12B, a model about six times bigger. ## Decision models Two new variants, d1-2B and d1-Omni-2B, answer several questions about one input in a single forward pass. They return calibrated probabilities. ## Fine-tuning Every model in the family can be fine-tuned on a single GPU. That, plus the small footprint, is what makes the release interesting for always-on local agents. Omar Sanseviero-style hype aside (@omarsar0 calls it huge potential), the pitch is simple: a phone-sized model that sees, listens and talks back, and may be enough for many agent jobs that people currently hand to much larger models. Whether the benchmark gaps hold up outside Liquid's own test setups is the thing to watch. ## Questions this post answers ### What is Liquid AI's LFM3 model family and what sizes does it come in? LFM3 is Liquid AI's new generation of Liquid Foundation Models, built on a tri-hybrid architecture with new multimodal encoders. The family consists of roughly 2B-parameter open-weight models covering text, vision, audio, omni, and decision tasks, all sharing a single backbone, making it a compact multimodal lineup rather than one giant model. _Developers evaluating small open-weight models for agent workloads can follow releases like this on daily.dev._ ### How does LFM3-1.7B perform against larger models like Qwen3.5-9B on agent benchmarks? LFM3-1.7B scored 38.4 on WebArena-Verified compared to 28.9 for Qwen3.5-9B, a model more than five times its size, according to Liquid AI's own evaluation. It also supports a 256K context window and is trained to work in agent harnesses such as Hermes Agent and Pi. The result has not been independently verified outside Liquid's own testing. _Anyone comparing small versus large models for agent tasks can track independent verification of claims like this via daily.dev._ ### Can LFM3 models be fine-tuned on consumer hardware? Yes, LFM3 models can be fine-tuned on a single GPU, according to Liquid AI. Combined with their roughly 2B parameter size and 256K context window, this makes the LFM3-1.7B variant suited for always-on local agents that run on personal hardware without needing a cloud bill. _Teams weighing local versus cloud-hosted agents can keep up with small-model releases like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 44 comments across hackernews (as of 2026-10-28). **TL;DR:** The release is met with strong interest, especially for local/on-device agentic and classification use cases, with the team actively answering technical questions; a few commenters report tooling friction (MLX loading errors, quantization quality issues) and question whether tiny models are suited for coding tasks. **Sentiment:** 60% positive · 30% mixed · 10% skeptical **The case for** - People highlight the small decision models (d1-2B) as strong for calibrated local classification even on minimal RAM. - Several note the 1.7B model performs well relative to larger models like Qwen3.5-4B/9B for agentic tasks given its size. - Commenters already using prior LFM versions for email/metadata annotation are excited to upgrade and see fine-tuning guides as lowering the barrier. - OCR and audio transcription + action-item extraction workflows are seen as promising use cases. **The pushback** - Some report the Omni and Audio MLX variants fail to load in tools like LM Studio due to safetensors read errors. - One commenter found the VL-2B model answered a basic text-only question incorrectly, suspecting quantization issues. - Skepticism that a model this small is a realistic fit for coding tasks compared to its agentic/automation strengths. - One user noted their home-assistant use case found an earlier 2.5 version 'too dumb,' raising doubts about whether 1.7B will be sufficient. **By community** - hackernews (positive): Enthusiastic, technical discussion dominated by practical use cases (classification, OCR, transcription, fine-tuning) with the team actively responding to questions, tempered by a few tooling and quality gripes. **Open questions** - How does LFM3-Omni-2B compare to LFM3-VL-2B specifically on vision tasks, and when should one be chosen over the other? - How many labeled examples are needed to fine-tune for specific tasks like audio-to-action-items? - How are confidence/calibration levels computed in the decision models (e.g., via logprobs)? **Highlights** > Depends on what you care about. If you want to do local classification/image processing, I think the d1-2B is amazing, it can run on 1GB of unified RAM. Now that it's released and I have weekends off again, I'm going to play with the VL-2B model and the d1-2B and see if I can prompt it to control a small Lego robot arm like this https://bricksrl.github.io/ProjectPage/ If you like agents, I think the 1.7B is very strong for its size and can be used to automate small stuff like classifying emails, cleaning your downloads folder, etc. It can also be fine-tuned for specific use cases with SFT/DPO/GRPO. The blog post talks about fine-tuning for wind-turbine classification. If there's things you want to automate, but GPT-6-Luna is not strong enough, and bigger models are too expensive, you might fine-tune one of the LFM3 models and use it locally for repetitive tasks. > — [samstevens on hackernews · 1 comments](https://news.ycombinator.com/item?id=50263207) > Working on the app with on-device transcription. Is interesting how the live mode is restarting the text every 30 sec, wonder how it can impact a thing that may be agreed over several minutes. Say I would like to get action items out of an hour long recording, am I supposed to go straight from the audio with LFM3-Audio-2B? Or is it transcribe first and give the text to LFM3-1.7B? What if I want to fine-tune the action items routine. How many labelled examples do you estimate such task may need? > — [maxiR8 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50266559) > It all depends on your hardware and throughput requirements. Suppose 1. I were building for my work laptop (M4 Max, 36GB RAM) 2. I had a batched use cases (transcribing + making action items for all meetings in a day) I would use the LFM3-Audio-2B for ASR (audio -> text transcript) and then LFM3-1.7B for extracting particular action items and adding them to my tool of choice (Linear, Jira, Notes app, etc). If you can fine-tune (100+ labeled samples, and a GPU), you likely can fine-tune LFM3-Audio-2B for audio -> action items in a single step. > — [samstevens on hackernews](https://news.ycombinator.com/item?id=50266953) > Just tried both the Omni model and the text only model, both as mlx versions, in lm studio omlx and rapid mlx, none of them could actually fire the model up, kept saying safetensors cant be read (yes I realize Claude should hear this not this thread). > — [dchuk on hackernews · 1 comments](https://news.ycombinator.com/item?id=50264925) > Surprisingly (or maybe not?), LFM3-VL-2B gave me better results than LFM3-1.7B on text-only prompts (general knowledge, chatting) through liquid playground. As an example, when asked “what does a cow say?”, LFM3-1.7B said that a cow can say “Oink”, and when I asked the same question to LFM3-VL-3B, it (correctly) said “moo”. I assume this is lower quant problem, though: Q8 versions I’ve played with through llama both give the correct “moo” answer. > — [lubitelpospat on hackernews · 2 comments](https://news.ycombinator.com/item?id=50265767) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50262229) · 105 points · 44 comments ## Similar posts on daily.dev - [Accelerating vision-language models with LFM2.5-VL-DSpark](https://daily.dev/posts/accelerating-vision-language-models-with-lfm2-5-vl-dspark-fkfjk19h6) · Hugging Face · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#multimodal](https://daily.dev/tags/multimodal) [View this post on daily.dev](https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Liquid AI announces LFM3, a new generation of Liquid Foundation Models","url":"https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh"},"datePublished":"2026-10-27T17:16:29.435Z","dateModified":"2026-10-28T10:20:09.814Z","description":"Liquid AI announced LFM3, the next generation of its Liquid Foundation Models. The company says the models use a new tri-hybrid architecture and new multimodal...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,multimodal","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Liquid AI announces LFM3, a new generation of Liquid Foundation Models"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/liquid-ai-announces-lfm3-a-new-generation-of-liquid-foundation-models-xbp2peueh#faq","mainEntity":[{"@type":"Question","name":"What is Liquid AI's LFM3 model family and what sizes does it come in?","acceptedAnswer":{"@type":"Answer","text":"LFM3 is Liquid AI's new generation of Liquid Foundation Models, built on a tri-hybrid architecture with new multimodal encoders. The family consists of roughly 2B-parameter open-weight models covering text, vision, audio, omni, and decision tasks, all sharing a single backbone, making it a compact multimodal lineup rather than one giant model. Developers evaluating small open-weight models for agent workloads can follow releases like this on daily.dev."}},{"@type":"Question","name":"How does LFM3-1.7B perform against larger models like Qwen3.5-9B on agent benchmarks?","acceptedAnswer":{"@type":"Answer","text":"LFM3-1.7B scored 38.4 on WebArena-Verified compared to 28.9 for Qwen3.5-9B, a model more than five times its size, according to Liquid AI's own evaluation. It also supports a 256K context window and is trained to work in agent harnesses such as Hermes Agent and Pi. The result has not been independently verified outside Liquid's own testing. Anyone comparing small versus large models for agent tasks can track independent verification of claims like this via daily.dev."}},{"@type":"Question","name":"Can LFM3 models be fine-tuned on consumer hardware?","acceptedAnswer":{"@type":"Answer","text":"Yes, LFM3 models can be fine-tuned on a single GPU, according to Liquid AI. Combined with their roughly 2B parameter size and 256K context window, this makes the LFM3-1.7B variant suited for always-on local agents that run on personal hardware without needing a cloud bill. Teams weighing local versus cloud-hosted agents can keep up with small-model releases like this on daily.dev."}}]} ```