Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHacker News AI→ original

Liquid AI: the small LFM3 model solved 64% of agent tasks on Mac, Qwen3.5-2B managed 48%

Liquid AI has released the weights for LFM3, a family of six models of roughly 2 billion parameters each that run on phones and laptops without the cloud. In MacPaw's test of 65 everyday agent tasks in real macOS, the text-only LFM3-1.7B (1.7 billion parameters) solved 64% of tasks against 48% for Qwen3.5-2B, and used 58% fewer tokens. For local agents, small models are therefore worth trying on your own tasks.

AI-processed from Hacker News AI; edited by Hamidun News
Liquid AI: the small LFM3 model solved 64% of agent tasks on Mac, Qwen3.5-2B managed 48%
Source: Hacker News AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHacker News AI ↗
Editor-in-Chief
Key metric #1Got stuck in fewer than 1% of attempts, versus 15% for Qwen3.5-2B
128K-token context cache1.6 GB versus 4.3 GB for Qwen3.5-4B
Key metric #3On-device decoding 1.8–2.8 times faster than the closest competitor

An agent that sorts your email, manages your calendar and navigates apps usually requires a large cloud model and a fee for every token. With the release of LFM3, Liquid AI argues that a model of about 2 billion parameters, running directly on a laptop or phone, is enough for many of these tasks. MacPaw's test offers a telling result: across 65 everyday agent tasks in real macOS, the text-only LFM3-1.7B (the 1.7 billion parameter version) completed 64% of tasks, while Qwen3.5-2B completed 48%. LFM3 also used 58% fewer tokens and got stuck in fewer than 1% of attempts, compared with 15% for its rival.

What's in the family

The first LFM3 release includes six models, all derived from a single base model, LFM3-1.7B-Base. Four accept different input types: text, images and video, audio, and everything at once (an omni model). The other two are "decision models," specialized versions built on the vision and omni models. Weights for both the base and fine-tuned versions are open.

Liquid's idea is not to run one large general-purpose model, but to pick a suitable narrow one, fine-tune it for your task and run it locally. For the text version, the company describes short fine-tuning on a single GPU. Because tokens on your own device are free, dozens of agents can run in the background around the clock. To demonstrate this, Liquid launched more than 300,000 LFM3-1.7B agents in parallel to process the Enron email archive.

How the model is built and why it is fast

The architecture is hybrid: of its 24 layers, 13 are short convolutions (cheap and fast), 5 are sliding-window attention with a 256-token window, and 6 are full global attention. An automated search selected this scheme by measuring speed across 14 combinations of devices and backends: Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, cloud Arm and x86 processors, GPUs, and Qualcomm Hexagon NPUs. The previous generation, LFM2, was tested on the processors of just one phone and one laptop.

The practical gain shows up with long context. A full 128K-token context takes 1.6 GB of cache, versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. In llama.cpp tests with Q4_0 quantization and a 32K context, the model decodes 1.8–2.8 times faster than the closest competitor on all three devices, and it uses the least memory. On a single NVIDIA B300 with SGLang, it produces 3.6 billion tokens per day at 1,024 concurrent requests.

Training used 22 trillion tokens. The last 3.5 trillion came in a stage on curated data with a large share of agent traces and tool calls, which also extended the context window to 256K. The model then went through four post-training stages, including distilling six narrow specialists (code, math, tools, long context, agentic work, and enterprise scenarios) into a single model.

Vision, audio and honest limitations

LFM3-VL-2B is the first in the lineup to accept video: you can ask questions about a lecture recording or a camera clip. On 62 vision benchmarks it beats models of its size by at least 7.4 points, matches Gemma 4 12B (which is more than six times larger), outperforms Gemini 3.

5 Flash Lite, and trails GPT-5.6 Luna by less than 2 points. Its own 164-million-parameter vision encoder requires 2.

2 times less compute than the Qwen3.5-2B encoder, with the same average score of 55.9% across 13 benchmarks.

It also uses three times less compute than SigLIP2 in Liquid's earlier models. The audio model accepts text or sound and replies with text, speech, or both; it runs locally on laptops and edge devices.

There are caveats. In agentic programming and knowledge tasks, larger models still win: here, size matters most. Adding search nearly doubles the score on Humanity's Last Exam, but does not close the gap. Most of the comparisons are published by Liquid itself, and MacPaw's test is the one that looks independent.

For developers, the takeaway is this. If you need local agents, document parsing, or a voice interface without sending data out, the 2B class is now a viable option. The weights are open, fine-tuning is cheap, and AMD, Qualcomm and Ambarella have announced support. It is worth running the model on your own tasks to see which of them really require a large cloud model.

Story

Key Questions Answered

What's in the family?

The first LFM3 release includes six models, all derived from a single base model, LFM3-1.7B-Base. Four accept different input types: text, images and video, audio, and everything at once (an omni model). The other two are "decision models," specialized versions built on the vision and omni models. Weights for both the base and fine-tuned versions are open. Liquid's idea is not to run one large general-purpose model, but to pick a suitable narrow one, fine-tune it for your task and run it locally.

How the model is built and why it is fast?

The architecture is hybrid: of its 24 layers, 13 are short convolutions (cheap and fast), 5 are sliding-window attention with a 256-token window, and 6 are full global attention. An automated search selected this scheme by measuring speed across 14 combinations of devices and backends: Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, cloud Arm and x86 processors, GPUs, and Qualcomm Hexagon NPUs.

What are the key details on Vision, audio and honest limitations?

LFM3-VL-2B is the first in the lineup to accept video: you can ask questions about a lecture recording or a camera clip. On 62 vision benchmarks it beats models of its size by at least 7.4 points, matches Gemma 4 12B (which is more than six times larger), outperforms Gemini 3.5 Flash Lite, and trails GPT-5.6 Luna by less than 2 points. Its own 164-million-parameter vision encoder requires 2.2 times less compute than the Qwen3.5-2B encoder, with the same average score of 55.9% across 13 benchmarks.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…