On-device structured extraction
Convert local text into a known schema without sending every input to a cloud model.
Independent tool overview
LFM2.5-350M at a glance
LFM2.5-350M is Liquid AI's 350-million-parameter text model for fast on-device extraction, structured output and tool calling across CPUs, GPUs and NPUs.
Visit the official LFM2.5-350M site ↗
- Parameters
- 350 million
- Context length
- 32,768 tokens
- Checkpoint size
- About 714 MB
- Languages
- 9
- Knowledge cutoff
- Mid-2024
- License
- LFM Open License v1.0
Overview
What LFM2.5-350M is
LFM2.5-350M is built for narrow edge workloads where a large general-purpose model would be too slow, expensive or memory-hungry. The official checkpoint is about 714 MB, and optimized builds can run below 1 GB of memory on phones, laptops, single-board computers and embedded devices.
Liquid AI recommends the instruction-tuned model for data extraction, structured outputs and tool use. Its small size makes local privacy, offline operation and high-volume processing practical, but it should be grounded in application data rather than trusted for broad world knowledge.
The model card explicitly advises against knowledge-intensive tasks and programming, while Liquid's launch post also excludes math and creative writing. A production design should give it a constrained schema, a small tool set and a larger-model or human fallback when the request leaves that lane.
Use cases
Who LFM2.5-350M is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Small tool-calling agents
Route a bounded request to a short list of functions on phones, laptops or embedded systems.
High-volume narrow processing
Run a well-defined classification or extraction step cheaply at the edge or with high GPU concurrency.
Capabilities
Core LFM2.5-350M features
Compact hybrid architecture
Uses 350 million parameters across convolution and grouped-query-attention blocks for efficient inference.
Function calling
Supports tool definitions and structured function-call output for constrained agent workflows.
Structured output
Is post-trained for extraction and instruction following rather than broad open-ended reasoning.
Nine-language support
Lists English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese and Spanish.
Broad runtime support
Ships in formats or integrations for Transformers, GGUF/llama.cpp, ONNX, MLX, OpenVINO, vLLM and SGLang.
Base and instruction checkpoints
Provides a base model for customization and a post-trained model for direct instruction-following use.
Process
How the LFM2.5-350M workflow works
Step 1
Define one bounded job
Choose a task with a fixed schema, known inputs and a clear test for success.
Step 2
Select the deployment format
Match GGUF, ONNX, MLX, OpenVINO or the native checkpoint to the target hardware and runtime.
Step 3
Evaluate and fine-tune
Measure extraction and tool-call accuracy on real examples, then customize if prompting alone is not reliable.
Step 4
Add validation and fallback
Validate every structured result and escalate low-confidence, out-of-scope or high-impact requests.
Cost
LFM2.5-350M pricing and free plan
The weights are downloadable under Liquid AI's custom LFM Open License v1.0. Free commercial use has a $10 million annual-revenue threshold; larger companies need a separate paid license.
Research and qualifying nonprofit use
$0 license fee
The license permits research and qualifying nonprofit non-commercial use without the revenue threshold.
- Self-hosting costs still apply
- Retain required notices
- Review the complete license
- No hosted service included
Commercial use under $10M revenue
$0 license fee
Legal entities below $10 million in annual revenue can use the model commercially under the published terms.
- Revenue is evaluated at the legal-entity level
- Redistribution requires attribution and the license
- Modified files must be marked
- Infrastructure costs are separate
Commercial use above $10M revenue
Contact Liquid AI
The free commercial grant ends above the threshold and a separate license is required.
- Commercial terms are not publicly priced
- Contact Liquid AI before deployment
- Applies to derivatives as described in the license
- Seek counsel for material commercial use
Pricing checked . Check current pricing at the source ↗
Assessment
LFM2.5-350M strengths and limitations
Where it stands out
- Small enough for practical deployment on phones, laptops and inexpensive edge devices
- Fast CPU and accelerator inference can make offline interaction responsive
- Tool use and structured extraction are unusually capable for the parameter count
- Multiple official formats reduce porting work across hardware families
- Local execution can reduce network latency and limit raw-data exposure
- Base checkpoint supports task-specific fine-tuning
What to consider
- Liquid AI does not recommend the model for knowledge-intensive tasks or programming
- Math, broad reasoning and creative writing are outside its strongest use cases
- The knowledge cutoff is mid-2024, so current facts require a trusted tool or data source
- Benchmark claims come from defined test setups and may not predict a specific device or workflow
- A 32K context window does not mean a 350M model can reliably reason over every long input
- The custom license is not Apache 2.0 and restricts free commercial use above $10 million in annual revenue
- Tool calls must be validated and permissioned; a small local model can still choose the wrong function
- Quantization can reduce memory and speed up inference while changing accuracy
Compare
LFM2.5-350M alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Qwen3.5 Small
A family of compact open models for teams willing to trade a larger footprint for broader capability.
Explore Qwen3.5 Small →Consumer
Gemma 4
A more capable open-weight family for tasks that exceed a narrowly scoped 350M edge model.
Explore Gemma 4 →Miscellaneous
DiffusionGemma
An experimental larger local model focused on very fast generation on dedicated GPUs.
Explore DiffusionGemma →Questions
LFM2.5-350M FAQs
What is LFM2.5-350M good for?
Liquid AI recommends it for data extraction, structured outputs and tool use, especially when the model needs to run locally or at high volume.
Can LFM2.5-350M run on a phone?
Yes. Liquid publishes mobile and edge deployment paths, and optimized versions have been demonstrated on iPhones, Android phones and other constrained devices.
How much memory does it need?
The official materials say it can run under 1 GB, with actual peak memory depending on hardware, runtime, precision, context length and quantization. The native checkpoint is roughly 714 MB.
Is LFM2.5-350M open source?
It is an open-weights model under the custom LFM Open License v1.0, not an OSI-standard open-source license. Commercial use above the license's $10 million annual-revenue threshold requires a separate agreement.
Does it support function calling?
Yes. The model card documents tool definitions, function-call output, tool responses and final answers. Applications should still validate every call and restrict permissions.
Should I use it as a general chatbot?
Only for a narrow, grounded assistant. Its size is an advantage for edge deployment, but Liquid advises against knowledge-intensive, programming, math and creative-writing workloads.
Bottom line
Our LFM2.5-350M verdict
LFM2.5-350M is a purpose-built edge component, not a miniature frontier chatbot. It is attractive when extraction or tool routing must happen locally with very little memory; its value depends on a tight task definition, evaluation on real data and a fallback for anything broader.
Visit LFM2.5-350M website ↗Similar Consumer Tools
View all Consumer AI tools →
Trinity-Large-Thinking
Trinity-Large-Thinking - Arcee AI's new open-weight frontier reasoning model for long-horizon agents
GLM-5.1
GLM-5.1 — Z AI's new open-source flagship model built for long-horizon agentic coding
Qwen3.5-Omni
Qwen 3.5 Omni - Alibaba's native omnimodal model with text, image, audio, and video understanding across 113 languages
Muse Spark
Muse Spark - Meta Superintelligence Labs' new multimodal reasoning model
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.