LFM 2.5: Architecture, Benchmarks and Why It Matters for Edge AI

All About the New LFM 2.5: Architecture, Benchmarks, and Why It Matters for Edge AI

Running capable AI without a cloud bill is no longer a fantasy. LFM 2.5 is Liquid AI's newest family of on-device foundation models, and its flagship agent, LFM2.5-2.6B, fits in under 2.5 GB of memory, runs at roughly 30 tokens per second on phone-class hardware, and beats or matches models several times its size on selected benchmarks. That makes private, offline AI genuinely practical. Keep in mind the headline figures are Liquid AI's own vendor benchmarks, so treat them as claims until independent runs confirm them.

This guide explains what LFM 2.5 is, how its architecture works, how the variants compare to Gemma, Qwen, and DeepSeek, and how to deploy it. If you also lean on cloud AI subscriptions like Claude, ChatGPT, or Gemini for heavier jobs, there is a smarter way to pay for those too, which we will cover along the way.

Building agents locally but still paying for cloud AI on the side? Bleap charges 0% FX fees on your USD subscriptions and a flat 20% cashback on Claude, ChatGPT, and Gemini renewals, with no subscription of its own. (The 20% cashback applies to those three tools only.) Get the Bleap card →

1. What Is LFM 2.5? Overview and Architecture

LFM 2.5 is the latest generation of Liquid Foundation Models, built by Liquid AI, a company founded by former MIT computer scientists. The family targets one goal: powerful intelligence that runs on the device in front of you, not in a data center.

The core innovation is a hybrid backbone. LFM2.5-2.6B is a 2.69B-parameter model with 30 layers (22 double-gated short convolution blocks and 8 GQA layers), a 128K vocabulary, and a 131,072-token context window. Replacing most attention layers with short convolutions cuts memory use and speeds up inference on ordinary CPUs.

How the Mixture of Experts Design Works

The flagship efficiency model layers a Mixture of Experts (MoE) design on top of that backbone. LFM2.5-8B-A1B uses a sparse MoE design that activates 1.5B of 8.3B total parameters per forward pass, which keeps each generated token cheap to compute. Each MoE layer has 32 experts and selects the top-k=4 experts per token with a normalized sigmoid router and adaptive routing biases for load balancing. The result is roughly 4B-class quality at close to 1.5B-class decode cost, which standard transformer-only models cannot match on constrained hardware.

2. LFM 2.5 Model Variants and Sizes

LFM 2.5 is not one model but a spread of sizes and specialisations.

LFM2.5-230M

The smallest checkpoint. Liquid AI shows LFM2.5-230M scoring 43.26 on the BFCLv3 tool-use benchmark, dominating IBM's Granite 4.0-350M (39.58) and outpacing larger 1-billion-parameter models like Google's Gemma 3 1B IT (16.61). It is built for structured tool calls and data extraction.

LFM2.5-1.2B Family

This is a comprehensive release with Base, Instruct, Japanese, Vision-Language, and Audio-Language models. Liquid AI extended pretraining from 10T to 28T tokens and scaled up post-training with reinforcement learning.

LFM2.5-2.6B

The general-purpose, phone-ready agent. It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots.

LFM2.5-8B-A1B (MoE)

The reasoning flagship. Unlike its predecessor, LFM2.5-8B-A1B is a reasoning-only model that produces an explicit chain of thought before its final answer.

3. On-Device and Edge AI Capabilities

The whole point of LFM 2.5 is running without the cloud or a GPU.

Running LFM 2.5 on Raspberry Pi and Mobile Hardware

The numbers are the practical signal. LFM2.5-230M runs at 213 tokens per second on a Galaxy S25 Ultra and 42 tokens per second on a Pi 5. For the larger agent, Liquid AI measured 220 tokens per second decoding on an Apple M5 Max, 113 tokens per second on an AMD Ryzen AI Max+ 395, and around 30 tokens per second on a phone, all under 2.5 GB of memory.

Privacy and Offline AI Advantages

Because inference stays local, data never leaves the device and the marginal cost of each run is near zero. That matters for healthcare, legal, and finance workloads, air-gapped environments, and anywhere a cloud API bill would balloon at scale.

4. LFM 2.5 Benchmarks and Performance Results

Liquid AI benchmarks against Gemma 4 and Qwen 3.5 variants using its own evaluation suite.

Text and Reasoning Benchmarks

For the reasoning model, the gains over its predecessor are large. LFM2.5-8B-A1B improves across the board: the AA-Omniscience Non-Hallucination Rate jumped from 7.46 to 63.47 and IFEval rose from 79.44 to 91.84.

Math and Agentic Task Benchmarks

MATH500 climbed from 74.80 to 88.76 and Tau² Telecom rose from 13.60 to 88.07. For the smaller agent, LFM2.5-2.6B leads on every instruction-following benchmark and nearly every tool-use benchmark, trailing only Qwen3.5-9B on BFCLv4, and outperforms both Gemma models across the board on agentic tasks.

Vision and Audio Benchmarks

The multimodal checkpoints hold their own. LFM2.5-VL-1.6B shows stronger instruction-following performance across both vision and text instruction benchmarks, making it a good choice for multimodal on-edge applications.

Inference Speed Benchmarks

Speed is the standout. LFM2.5-2.6B is the fastest model in its size class, reaching almost 15K output tokens per second at high concurrency, roughly 1.3B tokens per day on a single H100.

Prototyping agents locally, but renting cloud models for the hard parts? Pay those USD subscriptions with Bleap for 0% FX fees, plus a flat 20% cashback on Claude, ChatGPT, and Gemini. Self-custodial Mastercard, no monthly fee. Get the Bleap card →

5. LFM 2.5 vs. Competitors: Head-to-Head Comparisons

LFM 2.5 vs. DeepSeek-V4-Flash

DeepSeek-V4-Flash is a strong, cloud-oriented agent model. LFM 2.5's advantage is deployability: it runs on a phone or Raspberry Pi with no GPU, where a Flash-class model expects server hardware. If your workload lives on the edge, LFM 2.5 is the clear fit.

LFM 2.5 vs. Google Gemma

Liquid AI compares directly against Gemma 4 variants. Evaluated against gemma-4-E2B-it (5.1B) and gemma-4-E4B-it (8B) among others, LFM2.5-2.6B leads every instruction-following benchmark and every tool-use benchmark except BFCLv4. Gemma keeps an edge on coding.

LFM 2.5 vs. Alibaba Qwen

Qwen 3.5 is the closest rival on capability. On agentic tasks LFM2.5-2.6B trades closely with the Qwen models, leads on AA-Omniscience-Public, trails only Qwen3.5-9B on math, and coding is the one area where the larger models keep an edge. The active-parameter efficiency is where LFM 2.5 pulls ahead on cost per token.

The takeaway: LFM 2.5 delivers competitive accuracy at drastically lower compute, occupying a niche larger models cannot reach.

6. Agentic AI Design Philosophy

LFM 2.5 is built for agents first, chat second.

Tool Use and Agentic Workflows

The models support native function calling and multi-step planning. The MoE design matters here: fewer active parameters mean faster iteration through agentic loops. Zero per-token cost enables continuous background agent parallelism, a pattern that metered cloud APIs structurally prevent, so agents can be massively parallelized on local hardware at no marginal cost.

Practical Agentic Use Cases

Practical builds include on-device assistants, offline document triage over 128K inputs, form and invoice extraction, robotics command parsing, and background agents that run continuously without per-token cost. Be honest about scope, though: Liquid AI explicitly does not recommend the model for agentic coding or knowledge-heavy tasks.

7. Multimodal Capabilities: Vision and Audio

LFM 2.5 is a multimodal family, not a text-only model.

LFM 2.5 Vision-Language Model

The vision checkpoint handles images, charts, and documents alongside text, and the 1.6B VL variant is tuned for strong instruction-following on edge devices, which suits mobile document and image workflows.

LFM 2.5 Audio-Language Model

The audio-language model shipped as part of the 1.2B release, covering speech understanding and audio question answering. Potential applications include voice assistants, meeting summarisation, and accessibility tools, all running offline for privacy.

8. What Changed from LFM 2.0 to LFM 2.5

The jump is substantial. Liquid AI doubled the vocabulary to 128K by extending the tokenizer in place rather than retraining from scratch, to better support non-Latin scripts. Training scaled up sharply, with the 8B MoE trained on 38T tokens plus large-scale RL. LFM 2.5 also added a 128K context window, reasoning, and multimodal Vision and Audio models not present in 2.0.

9. Licensing and Commercial Availability

What the License Allows

The models ship under Liquid's own license. The base model and the agentic post-trained variant are both on Hugging Face under the LFM Open License v1.0, which allows commercial use. That said, the model ships under Liquid's own license rather than a standard permissive one, so anyone with commercial plans should read the terms before building on it.

Where to Access LFM 2.5

You can download weights and model cards from Hugging Face, or access the models via OpenRouter and Liquid AI's own API for rapid prototyping.

10. How to Get Started and Deploy LFM 2.5

Quick-Start via API

Create an account with OpenRouter or Liquid AI, obtain an API key, select your variant via the endpoint, and send a test request. This is the fastest path to a working prototype.

Local Setup via HuggingFace

Install your inference stack, download the weights for your chosen variant from the Hugging Face model card, and run inference locally. Day-one support spans llama.cpp, MLX, vLLM, and SGLang, with open base and post-trained weights.

Edge Deployment

For Raspberry Pi and mobile, use quantised formats such as GGUF with INT4/INT8. For the MoE flagship on GPU, size the GPU for the full ~8B of weights (≈17 GB BF16 plus KV cache) even though only ~1B is active per token, targeting a GPU with at least 24 GB VRAM.

11. Real-World Value: Business and Personal Applications

For Individual Developers and Hobbyists

Build offline AI apps on consumer hardware with no recurring API cost, experimenting freely via Hugging Face.

For Enterprises

Private, on-premises deployment eliminates data-sharing risk, cuts inference costs versus GPT-4-class APIs at scale, and enables vertical fine-tuning on proprietary data.

For Edge and IoT Industries

Manufacturing quality control with the vision model, smart-home and robotics agents with the 8B-A1B, and diagnostics support in connectivity-limited regions are all in reach.

Shipping an AI product but still burning cash on cloud model subscriptions? Bleap gives you 0% FX fees on USD billing and a flat 20% cashback on Claude, ChatGPT, and Gemini, with a self-custodial Mastercard and no monthly subscription. Get the Bleap card →

Frequently Asked Questions About LFM 2.5

What hardware does LFM 2.5 require, and can it really run on a Raspberry Pi?

Yes. The smaller variants run on constrained hardware with quantisation. LFM2.5-230M runs at 42 tokens per second on a Pi 5, while the 2.6B agent runs under 2.5 GB of memory, including around 30 tokens per second on a phone.

How does LFM 2.5 compare to DeepSeek and Qwen at the same parameter count?

Against Qwen, LFM2.5-2.6B leads on every instruction-following benchmark and nearly every tool-use benchmark, trailing only Qwen3.5-9B on BFCLv4. Its active-parameter efficiency means far lower compute per token than dense rivals or larger cloud models like DeepSeek-V4-Flash.

Is LFM 2.5 free to use commercially?

Largely, yes. The weights are on Hugging Face under the LFM Open License v1.0, which allows commercial use, but you should read the license terms before building a commercial product on it.

Where can I download or access LFM 2.5 models?

On Hugging Face for weights, and via OpenRouter or the Liquid AI API for hosted access.

Does LFM 2.5 support vision and audio, or just text?

It is multimodal. The release includes Base, Instruct, Japanese, Vision-Language, and Audio-Language models.

How fast is LFM 2.5 inference compared to similar-sized models?

Very fast for its class. LFM2.5-2.6B reaches almost 15K output tokens per second at high concurrency, thanks to the short-convolution backbone and sparse MoE routing.

Conclusion: Is LFM 2.5 the Edge AI Model to Watch?

LFM 2.5 earns its attention: competitive benchmarks, native multimodal support, and an architecture powerful enough for enterprise agents yet light enough for a Raspberry Pi. Developers prototyping agents, enterprises evaluating private AI, and edge hardware engineers should all take a serious look now, starting on Hugging Face or the API.

And whichever AI tools you run alongside it, pay smart. With Bleap you skip the FX fees on USD subscriptions, and on Claude, ChatGPT, and Gemini you earn 20% cashback on every renewal, with a self-custodial Mastercard and no subscription of its own.