LFM 2.5: Liquid AI's small models for iPhone and Mac, explained
· By Modelbase Team
LFM 2.5 is a family of small open AI models from Liquid AI, designed to run on phones, laptops and other devices rather than in a data center. The main model, LFM2.5-1.2B-Instruct, has about 1.2 billion parameters and fits in a 663 MB download, small enough for an iPhone. The family also includes a 350M model for older phones, a “Thinking” model that reasons step by step, and vision models that can look at photos. Liquid AI announced LFM 2.5 in January 2026, and it’s the model that comes built into Modelbase.
Who is Liquid AI, and what is an LFM?
Liquid AI is an AI company that builds its own model architecture instead of using the standard Transformer design alone. LFM stands for Liquid Foundation Model. Its second generation, LFM2, was described in a technical report in November 2025, and LFM 2.5 builds on it.
What makes the LFM 2.5 models different is the inside. LFM2.5-1.2B-Instruct is a hybrid model with 16 layers: 10 double-gated convolution blocks and 6 grouped-query attention blocks, and Liquid AI says it’s “designed for on-device deployment” (Liquid AI). In plain terms: most of the model is made of a lighter kind of layer than the attention layers of a standard language model, which suits small hardware.
What’s new in LFM 2.5
Liquid AI announced LFM 2.5 on 5 January 2026 (Liquid AI). Compared with LFM2:
- More training. Pretraining was extended from 10 trillion to 28 trillion tokens.
- More reinforcement learning after pretraining.
- New models: a 1.2B Base model for fine-tuning, 1.2B Instruct for chat, a Japanese chat model, a 1.6B vision-language model and a 1.5B audio model.
More sizes have followed, including the 350M model and the Thinking model that Modelbase offers.
LFM 2.5 models you can run on your iPhone and Mac
These are the LFM 2.5 models in Modelbase’s catalog, with their download sizes and the iPhones the app recommends them for:
| Model | What it’s for | Download | Recommended for |
|---|---|---|---|
| LFM 2.5 (350M) | Quick everyday chat, summaries and short drafts | 227 MB | iPhone 14 and older |
| LFM 2.5 (1.2B), included | Everyday assistant: writing, summaries, questions about documents | 663 MB | iPhone 15 and newer |
| LFM 2.5 Thinking (1.2B) | Reasons before answering: math and logic | 663 MB | iPhone 15 and newer |
| LFM 2.5 VL (450M) | Quick questions about photos | 381 MB | iPhone 14 and older |
| LFM 2.5 VL (1.6B) | Questions about photos, in more languages | 1.5 GB | iPhone 15 Pro and newer |
| LFM 2.5 VL (3B) | Liquid AI’s largest vision model: images, text in photos, documents | 2.39 GB | iPhone 15 Pro and newer |
All of them are 4-bit MLX versions, the format Modelbase runs with Apple’s MLX framework. On a Mac with Apple silicon they’re all small enough to run comfortably.
What LFM 2.5 is good at, and what it isn’t
Liquid AI is unusually direct about this. For LFM2.5-1.2B-Instruct it recommends agentic tasks, data extraction and RAG (answering from documents you provide), and it doesn’t recommend it for knowledge-intensive tasks and programming (Liquid AI).
In practice that means:
- Good: rewriting and polishing text, summarizing, pulling facts out of a PDF or an email, answering from a document, planning, short explanations, conversation.
- Not its strength: trivia and obscure facts (a 1.2B model can’t hold much world knowledge), and writing code.
That fits how Modelbase uses it. When you chat with a PDF, the app finds the relevant passages and the model answers from them; when you ask about the news, web search brings in sources. Both lean on what LFM 2.5 does well.
Languages: the 1.2B model supports English, Arabic, Chinese, French, German, Japanese, Korean and Spanish.
Context: Liquid AI lists a context length of 32,768 tokens for the 1.2B models. On a phone, the app uses a smaller window that fits the device’s memory; see what a token is.
LFM 2.5 Thinking
LFM2.5-1.2B-Thinking is a reasoning model of the same size: it writes out its thinking before it answers, which helps with math and logic problems. Liquid AI calls it a “general-purpose reasoning model” and gives it the same recommendations as the Instruct model (Liquid AI). In Modelbase it always reasons, so its thinking control offers Medium and High. Our guide to reasoning models explains what those levels do.
LFM 2.5 VL: the vision models
The VL (“vision-language”) models can look at a photo and answer questions about it: what a plant is, what a sign says, what’s on a receipt. Liquid AI says LFM2.5-VL-1.6B handles prompts in Arabic, Chinese, French, German, Japanese, Korean and Spanish as well as English (Liquid AI). See how to ask AI about a photo on your iPhone.
How to run LFM 2.5 on iPhone, iPad and Mac
Modelbase is free on the App Store for iPhone and iPad (iOS 18 or later) and Macs with Apple silicon (macOS 15 or later).
- Install Modelbase. LFM 2.5 (1.2B) is inside the app, so you can start chatting right away, offline, with nothing to download.
- Add more LFM 2.5 models. Tap the model’s name at the top of the chat, then Manage models > Local, and search for “LFM 2.5”. Tap Download on the ones you want.
- Switch models from the same model picker. Each model’s answers are signed with its name and “(Local)”.
On an older iPhone, start with LFM 2.5 (350M) and LFM 2.5 VL (450M). On an iPhone 15 Pro or newer, or any Apple silicon Mac, the whole family runs.
LFM 2.5 vs other small models
| Model | Maker | Parameters | Download in Modelbase |
|---|---|---|---|
| LFM 2.5 (1.2B) | Liquid AI | 1.2B | 663 MB |
| Qwen 3.5 (0.8B) | Qwen | 0.8B | 0.65 GB |
| Llama 3.2 (1B) | Meta | 1.2B | 0.71 GB |
| Gemma 3 (1B) | 1.0B | 0.77 GB | |
| Qwen 3.5 (2B) | Qwen | 2B | 1.75 GB |
At the same size, the choice comes down to the job: LFM 2.5 for documents and everyday assistance, Qwen 3.5 if you want one model that can also reason and see images, Llama 3.2 for plain multilingual dialogue. Our guides to Llama 3.2 on iPhone, Gemma vs Llama vs Qwen and which model to run on your iPhone go into more detail, and how to run an LLM locally covers the basics.
License: can you use LFM 2.5 commercially?
The LFM 2.5 models are open-weight under Liquid AI’s LFM Open License. Research and non-profit use is allowed without limits; companies can use them commercially for free while their annual revenue is under $10 million, and need a commercial license from Liquid AI above that (Liquid AI). For simply chatting with the model in an app, the license isn’t something you need to think about.
What’s next: LFM3
Liquid AI has since released LFM3: a 1.7B LFM3 model with a 256K-token context is listed on OpenRouter as of November 2026 (OpenRouter). LFM 2.5 remains the family in Modelbase’s catalog.
FAQ
What is LFM 2.5?
LFM 2.5 is a family of small open models from Liquid AI, announced in January 2026, built to run on devices. It includes 350M and 1.2B chat models, a 1.2B reasoning model and vision models from 450M to 3B.
Is LFM 2.5 good?
For its size, at the jobs Liquid AI designed it for: writing, summarizing, extracting information and answering from documents. Liquid AI doesn’t recommend it for knowledge-heavy questions or programming.
Can LFM 2.5 run on an iPhone?
Yes. LFM 2.5 (1.2B) is a 663 MB download and comes built into Modelbase; the 350M model (227 MB) suits iPhone 14 and older.
Does LFM 2.5 work offline?
Yes. Once a model is on your device, it runs with no connection. In Modelbase, LFM 2.5 (1.2B) works offline from the first launch.
What languages does LFM 2.5 support?
The 1.2B model supports English, Arabic, Chinese, French, German, Japanese, Korean and Spanish.
Is LFM 2.5 free?
Yes, the models are free to download and run, and Modelbase is free. Liquid AI’s license requires a commercial license only for companies with more than $10 million in annual revenue.
Try it on your iPhone, iPad or Mac
Modelbase is free on the App Store.