On-device application teams
Build assistants, extraction tools, and agents that need low latency, offline operation, or local data processing on mobile and embedded hardware.
Independent tool overview
LFM 2.5 at a glance
LFM2.5 is Liquid AI's expanding family of compact open-weight text, vision, audio, and retrieval models designed to run efficiently on phones, laptops, vehicles, and other edge hardware.
Visit the official LFM 2.5 site ↗
- Product type
- Open-weight edge AI model family
- Modalities
- Text, vision, audio, and retrieval
- Deployment
- Mobile, desktop, embedded, on-prem, or cloud
- Free commercial threshold
- Under $10M annual company revenue
- Reviewed
- August 29, 2026
Overview
What LFM 2.5 is
LFM2.5 is a model family rather than a consumer chatbot. Liquid AI launched it with compact 1.2B-parameter text, vision, audio, and Japanese variants, then expanded the family across smaller and larger text models, vision-language models, encoders, embeddings, and specialized checkpoints.
The central advantage is deployment efficiency. LFM2.5 models are distributed in formats for Transformers, GGUF and llama.cpp, MLX, ONNX, and other runtimes, making them practical for local inference on CPUs, GPUs, and NPUs. This can reduce cloud dependence, latency, and exposure of user data when an application truly runs on-device.
LFM2.5 is best for engineering teams that can evaluate, integrate, and monitor a small model for a defined task. It is not a drop-in replacement for a frontier cloud model across every workload, and the LFM Open License has a commercial revenue threshold that larger organizations must address before deployment.
Use cases
Who LFM 2.5 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Privacy-sensitive products
Keep inference on user-controlled hardware when the product architecture and chosen runtime do not send prompts to a cloud service.
Model customization
Fine-tune a compact base or instruction model for a narrow domain, language, or structured task.
Hardware-constrained systems
Deploy useful language, vision, or audio behavior in vehicles, laptops, phones, IoT devices, and other environments with limited compute or memory.
Local agents
Use instruction and tool-calling variants as components in applications that invoke approved functions without requiring a large hosted model.
Capabilities
Core LFM 2.5 features
Broad compact model family
Current LFM2.5 releases include text models from 230M through 2.6B parameters, an 8B mixture-of-experts model, vision-language variants, audio models, and retrieval-focused encoders.
On-device efficiency
The hybrid architecture combines convolution and attention blocks to target faster inference and lower memory use than conventional models of a similar size.
Instruction and tool use
Instruction-tuned checkpoints support chat, structured instruction following, and function calling for agent-like application workflows.
Multiple deployment formats
Liquid publishes native, GGUF, MLX, and ONNX checkpoints across much of the family, with support varying by model.
Quantization options
Developers can choose lower-bit GGUF, MLX, or ONNX variants to trade some model quality for smaller storage, memory, and compute requirements.
LEAP platform
Liquid's free LEAP platform provides model search, testing, fine-tuning tools, model bundling, and an Edge SDK for local deployment.
Fine-tuning without copyleft
The LFM license allows teams below its commercial threshold to keep their fine-tuned derivatives private while retaining required notices.
Specialized modalities
The family includes native vision-language and audio-language models rather than limiting developers to text-only inference.
Process
How the LFM 2.5 workflow works
Step 1
Define the constrained task
Start with a measurable use case, target hardware, response-time budget, memory ceiling, privacy requirement, and acceptable failure rate.
Step 2
Select a checkpoint
Compare the smallest text, vision, audio, or retrieval model that supports the task and confirm its context, languages, license, and runtime formats.
Step 3
Benchmark on target hardware
Test accuracy, latency, prefill speed, decode speed, memory, battery impact, and thermal behavior on the actual device rather than relying on vendor benchmarks alone.
Step 4
Specialize and package
Use prompting, retrieval, adapters, fine-tuning, and quantization only where evaluation shows they improve the target task, then create a deployment bundle.
Step 5
Integrate guardrails
Validate tool calls, constrain outputs, handle unsupported requests, and preserve a reliable fallback when the local model is uncertain.
Step 6
Monitor and update
Track real-world failures, regression-test new checkpoints, and revisit the license before company revenue or use conditions change.
Cost
LFM 2.5 pricing and free plan
Liquid AI makes its open models and LEAP's core features free to use. The LFM Open License permits commercial use at no charge while the user's company remains below $10 million in annual revenue; organizations at or above that threshold need a separate commercial license. Hardware, development, support, and any third-party hosting costs remain separate.
Open model use
Free below $10M annual revenue
Download, run, fine-tune, and deploy LFM2.5 under the LFM Open License.
- Commercial use is allowed below the annual revenue threshold
- Attribution and license notices must be retained when distributing
- Fine-tuned models can remain private
- Compute and deployment costs are not included
Research and qualified nonprofit use
Free
The license provides research and qualifying nonprofit allowances without the commercial revenue threshold.
- License terms still apply
- Review the formal definition of a qualified nonprofit
- Retain required notices when redistributing
LEAP
Free
Core model search, fine-tuning, bundling, and Edge SDK capabilities.
- No-cost core platform
- Model downloads
- Fine-tuning tools
- Model Bundling and Edge SDK
Enterprise
Contact sales
Commercial license and deployment support for companies at or above $10 million in annual revenue or teams needing a custom engagement.
- Commercial license beyond the threshold
- Custom support and scaling terms
- Pricing depends on the deployment
Pricing checked . Check current pricing at the source ↗
Assessment
LFM 2.5 strengths and limitations
Where it stands out
- Small checkpoints target real deployment constraints instead of maximizing parameter count
- Text, vision, audio, and retrieval variants cover more than a single chat use case
- Multiple runtime formats make local experimentation and packaging more accessible
- On-device execution can improve latency, offline availability, and privacy when implemented correctly
- LEAP's core tools and model downloads are free
- Teams below the revenue threshold can use and fine-tune the models commercially without a copyleft requirement
What to consider
- LFM2.5 is an engineering building block, not a polished assistant that nontechnical users can deploy by clicking once
- Compact models generally have less broad knowledge and reasoning capacity than much larger frontier systems
- The 1.2B Instruct model's published knowledge cutoff is mid-2024, and other variants can have different training details
- Benchmark results published by the model maker should be reproduced on the actual task and hardware before a buying or architecture decision
- Runtime and format support varies by checkpoint; not every model is available in every GGUF, MLX, or ONNX configuration
- The LFM Open License is not a standard Apache 2.0 license because free commercial rights end at the $10 million annual-revenue threshold
- Local inference does not automatically guarantee privacy if the surrounding app, telemetry, retrieval system, or tool calls still transmit data
- Teams must supply their own application layer, safety controls, monitoring, and usually task-specific evaluation data
Compare
LFM 2.5 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Gemma 4
Google's current open-weight family is a strong alternative when ecosystem breadth and a wider range of reasoning-oriented sizes matter.
Explore Gemma 4 →Coding
Meta Llama Models
Meta's Llama family offers a large deployment ecosystem and many third-party tools, fine-tunes, and hosting options.
Explore Meta Llama Models →Agents
Mistral AI
Mistral provides both open-weight and hosted model options for teams that may want to move between self-hosted and managed inference.
Explore Mistral AI →Questions
LFM 2.5 FAQs
What is LFM2.5?
LFM2.5 is Liquid AI's family of compact open-weight text, vision, audio, and retrieval models designed for efficient deployment on edge devices and conventional servers.
Is LFM2.5 open source?
Liquid calls the models open-weight. They can be downloaded, modified, and redistributed under the LFM Open License, but that license includes attribution obligations and a commercial revenue threshold, so it is not identical to a standard permissive software license.
Can I use LFM2.5 commercially for free?
Yes, while your company remains below $10 million in annual revenue and you comply with the LFM Open License. At or above the threshold, commercial use requires a separate license from Liquid AI.
What model sizes are available?
The current family includes compact 230M, 350M, 1.2B, and 2.6B text checkpoints, an 8B mixture-of-experts model, and separate vision, audio, embedding, and encoder variants. Check the live model library because the family continues to expand.
Can LFM2.5 run on a phone?
Yes. Liquid targets iOS, Android, and other edge hardware through LEAP and publishes quantized checkpoints for local runtimes. Actual speed and memory use depend on the chosen model, quantization, device, and app.
Does LFM2.5 support tool calling?
Selected instruction models support native function calling, but developers still need to validate tool calls and enforce permissions in the surrounding application.
How much does LEAP cost?
Liquid says LEAP's core model search, fine-tuning, bundling, and Edge SDK features are free. Enterprise support and licensing use custom terms.
Bottom line
Our LFM 2.5 verdict
LFM2.5 is a strong option when the hard requirement is useful AI within a strict memory, latency, connectivity, or privacy envelope. Its formats and expanding model family make it practical for serious edge evaluation, but teams should treat task accuracy, device performance, and the $10 million commercial threshold as first-class deployment requirements.
Visit LFM 2.5 website ↗Similar Consumer Tools
View all Consumer AI tools →
Alexa.com
Alexa.com - Amazon's AI assistant experience across voice, mobile, and web
Falcon H1R 7B
FalconH1R 7B - TII's small hybrid reasoner that outperforms rivals up to 7x its size on math and coding benchmarks while running at double the inference speed.
MiMo-V2-Flash
MiMo-V2-Flash - Xiaomi's powerful open-weights reasoning model
Gmail
Gmail - Google's email inbox, now infused with Gemini for AI-powered insights, actions, and improvements
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.