Mercedes-Benz Explores Local AI for Its MBUX Virtual Assistant

The next major shift in artificial intelligence may not come from models with hundreds of billions of parameters. It could come from smaller models that run directly on smartphones, computers and company-owned servers.

Liquid AI's new LFM3 family makes that possibility more concrete. Its flagship 1.7-billion-parameter language model reports stronger results than Qwen3.5-9B on selected browser automation and function-calling benchmarks.

The commercial opportunity is straightforward. Businesses could gain AI agents that operate software, process documents and handle voice commands without depending entirely on external AI infrastructure.

But smaller does not automatically mean better. LFM3 also reveals an important distinction between models that excel at specific business tasks and models designed for complex reasoning.

Liquid AI Introduces Six Models for Enterprise Automation

Released on October 27, LFM3 includes six models covering language, vision, audio and structured decision-making.

The family shares a 1.7-billion-parameter foundation, with specialized models reaching approximately 2 billion parameters.

Liquid AI says companies can customize these models using a single GPU and deploy them across local devices, private servers or cloud infrastructure.

The release also identifies Mercedes-Benz and Shopify as organizations involved in pre-release testing and evaluation.

This is an important development because enterprise AI adoption increasingly depends on deployment flexibility, operational costs and data control, not simply benchmark performance.

Smaller Models Are Becoming More Specialized

Traditional AI competition has focused heavily on model size. Larger models generally offer greater capabilities, particularly in knowledge-intensive tasks and complex reasoning.

LFM3 approaches the market differently.

Its architecture combines convolution layers with attention mechanisms designed to process information across different context windows. Liquid calls this a tri-hybrid architecture.

The shared foundation was pretrained on 22 trillion tokens, followed by training focused on tool usage, business workflows and agent capabilities.

According to Liquid, LFM3-1.7B improves upon its previous 2.6-billion-parameter model despite having approximately 35% fewer parameters.

The underlying business strategy is clear. Instead of building one model that attempts everything, Liquid is developing smaller models optimized for specific workloads.

Browser Automation Produces the Most Interesting Results

The strongest evidence for LFM3 comes from browser automation benchmarks.

Liquid reports the following results.

These are company-reported results, not independent benchmark verification.

The distinction matters.

LFM3 performs particularly well in selected browser automation and function-calling evaluations. However, Qwen3.5-9B remains stronger in several coding, knowledge and other agent-related evaluations.

Liquid itself acknowledges that larger models retain advantages in agentic coding and knowledge-intensive workloads.

For businesses, this suggests that model selection should begin with the actual task rather than parameter count.

Local Hardware Could Become a Competitive Advantage

One of LFM3's most interesting characteristics is its relatively small memory footprint.

Liquid says the text model can fit into approximately 1GB at four-bit precision. However, this represents model storage requirements, not the complete memory needed to run an application.

Its published hardware measurements provide a more realistic picture.

These measurements use four-bit Q4_0 models through llama.cpp, with a 32K-token input and one request at a time.

The Apple M5 Max result is particularly interesting. Generating 237.1 tokens per second demonstrates how smaller models can make sophisticated AI workloads practical on personal hardware.

However, sustained performance, battery consumption and thermal behavior still need evaluation for real-world applications.

Six Models Expand AI Beyond Text

LFM3 is not simply a language model release.

Its specialized models address several different enterprise requirements.

• LFM3-1.7B handles language generation and software tool usage.

• LFM3-VL-2B interprets images and video.

• LFM3-Audio-2B processes speech and text across 18 languages.

• LFM3-Omni-2B combines visual, audio and language inputs.

• d1-2B produces structured decisions from text and images.

• d1-Omni-2B extends structured decision-making to audio.

This creates opportunities beyond traditional chatbots.

A warehouse employee could describe damaged inventory verbally while capturing an image. A technician could photograph equipment information and generate a maintenance record. An internal assistant could classify customer requests without requiring employees to complete every form manually.

These are potential applications rather than confirmed production deployments.

Vision AI Could Improve Industrial Inspection

Liquid also demonstrates how specialization can improve performance.

The company fine-tuned LFM3-VL-2B using approximately 3,000 public wind-turbine images and several hours of training on an Nvidia A100.

Testing involved approximately 1,400 previously unseen images.

Classification accuracy increased from below 10% to 67%.

However, the detection F1 score reached only 0.11, compared with 0.003 before specialization.

This difference is important.

Recognizing a general defect category is not equivalent to reliably identifying and locating defects. Industrial customers still need human verification and further model improvement before considering autonomous inspection.

Voice AI Could Change Business Software Interaction

LFM3-Audio introduces another potentially valuable capability.

Liquid says its speech encoder can process incoming audio before the speaker finishes speaking. The company reports that the first response token can arrive as quickly as 50 milliseconds afterward.

That is a best-case measurement, not a guarantee of complete conversational response time.

For businesses, the bigger opportunity is reducing dependence on keyboards and conventional application interfaces.

Employees could speak requests, provide images and receive structured outputs.

LFM3-Omni takes this further by combining visual, audio and language processing in one model.

The Business Economics of Local AI Are Changing

Liquid offers both hosted API access and downloadable model weights.

Its listed hosted pricing includes:

Audio input and output carry additional charges.

For LFM3-1.7B, processing one million uncached input tokens and generating one million output tokens costs $0.28.

This makes experimentation relatively accessible.

However, API pricing is only one part of the financial equation. Companies must also consider application development, integration, monitoring and human review.

Private Deployment Comes With Licensing Requirements

Local deployment offers greater control over where inference happens.

It can also reduce dependence on external API providers and potentially limit the transfer of sensitive information.

Liquid states that its hosted models offer zero data retention and that submitted content is not retained or used for training.

Nevertheless, local inference does not automatically guarantee complete privacy. Connected applications can still transmit information through external search, messaging and other services.

There is also a licensing consideration.

LFM3-1.7B uses the LFM Open License v1.0. Under Liquid's published terms, businesses exceeding $10 million in annual revenue need a separate agreement for commercial use.

Hardware, electricity, maintenance and licensing therefore remain part of the deployment cost.

Competition Will Depend on Workloads Rather Than Model Size

LFM3 competes in a market where larger models still have advantages.

Qwen3.5-9B demonstrates stronger results in several reasoning and coding evaluations. Meanwhile, Liquid's approach emphasizes efficient execution, specialized training and flexible deployment.

Neither strategy makes the other obsolete.

The likely enterprise architecture will combine different models according to workload complexity.

A smaller model might classify incoming requests, extract invoice information or execute approved browser actions. A larger model could handle exceptions, complex reasoning and difficult coding tasks.

This layered approach could improve cost efficiency without sacrificing capability where it matters.

Small Businesses Could Benefit From Narrow AI Automation

For small and medium-sized businesses, LFM3's most relevant opportunity is not building an expensive AI platform.

It is automating one repetitive process.

Consider a company receiving hundreds of customer inquiries. A small model could classify messages, extract customer information and route requests to the appropriate department.

An ecommerce operation could experiment with document extraction, product-image interpretation and internal inventory workflows.

A service company could test voice-based maintenance reporting.

The important metrics are measurable outcomes:

• Employee time saved per task.

• Accuracy of extracted information.

• Percentage of requests processed successfully.

• Cost per completed workflow.

• Frequency of human intervention.

A small business should not deploy AI simply because a model is inexpensive. It should deploy AI when the resulting workflow costs less, operates faster or delivers better service.

The Next Enterprise AI Advantage May Be Smaller Models

Liquid AI's LFM3 release highlights a broader change in the AI industry.

The market is moving beyond the assumption that every useful application requires the largest available model.

Specialized smaller models can deliver meaningful advantages in automation, inference speed and deployment flexibility.

But benchmark victories do not automatically translate into business value. Real-world accuracy, licensing, infrastructure and operational reliability remain decisive.

For enterprises, the strongest opportunity is a hybrid AI architecture that uses small models for predictable tasks and larger models when deeper reasoning is necessary.

For smaller businesses, the message is even simpler.

Start with one expensive or repetitive workflow. Test whether a specialized model can handle it reliably. Measure the savings before expanding.

The companies that benefit most from AI may not be those that spend the most on the largest models. They may be those that understand exactly where smaller models can do useful work.

Mercedes-Benz Explores Local AI for Its MBUX Virtual Assistant

Liquid AI LFM3, LFM3-1.7B, local AI models, enterprise AI agents, small language models, AI browser automation, on-device AI, AI voice models, AI vision models, private AI deployment, AI business automation, Liquid AI pricing, AI model benchmarks, enterprise automation