The development of automated customer support systems is often associated with slow data processing. However, Liquid AI has introduced the D1 model, based on zero-token technology, which can instantly deliver structured results without generating redundant tokens. This review will test the model's performance using images of damaged goods and a JSON file of customer complaints to generate reports based on user requests.

How Liquid D1 Works

Liquid AI's D1 model evaluates all inputs in a single pass, enabling processed reports without the delays associated with text generation. The system can handle various media types, including images and text, simultaneously, distributing requests to the appropriate categories.

Liquid AI offers two versions of the model: 3 billion and 600 million parameters. Both versions function exclusively as decision-making engines and do not generate conversational text, relying on categorization nodes. This approach eliminates text generation-related delays. The model processes standard codes and images within a single pass, analyzing visual data directly. Furthermore, the system can handle a significant amount of background information, allowing for long documents to be sent with requests for deep routing. In the Semantic Tool benchmark, the D1 model achieved a score of 75%, ranking first in its weight class.

Technical Aspects and Installation

For testing, a 600 million parameter version of the model was chosen. Various file quantization options are available, affecting download size and RAM requirements. For this review, 8-bit quantization was selected, with a download size of 0.5 GB and RAM consumption between 2 and 4 GB. Additional project files were also required to enable visual capabilities. Downloading the necessary files is done using two commands.

The model's operational process involves the following stages:

1. Data Ingestion: The system accepts various input types, including text, code blocks, and media files, into a unified workspace.
2. Image Processing: A specialized module processes images, converting pixel data into embeddings that the model can understand.
3. Analysis: The core engine processes batches of text and images simultaneously, evaluating requests in parallel without latency.
4. Output Generation: The system bypasses the text generation phase, directly reading data from probability nodes and delivering structured responses instantaneously.

As a result, the model delivers solutions with zero output tokens in under 20 milliseconds. The model can be configured for various response types, including binary (yes/no) for quick verification or selection from a list of options. The system assigns a probability to each possible categorization and directs the query to the most suitable category. Additionally, the model can evaluate input data on a given scale, allowing for urgency ranking or rating assignment without complex prompts.

Practical Application

The D1 model can be used to automate return verification processes by analyzing product images to identify actual damage and approve valid claims in under a second. The system is also suitable for routing customer support requests, instantly sorting technical and billing inquiries and directing them to the appropriate departments. Furthermore, the model can serve as a protective mechanism for AI agents interacting with the command line, blocking malicious instructions before execution, thereby reducing computational costs and enhancing security.

Model Testing

After loading the model files (GGML), they were registered in the Olama service under the name liquid-d1. The input data consisted of images of damaged products (dishware, luggage) and corresponding customer messages in JSON format.

Example queries included:

• Binary Question: Determining the presence of physical damage in the photographs.
• Category Selection: Identifying the department responsible for handling the claim (fragile replacement department, hazardous appliances department, luggage logistics team).
• Damage Severity Assessment: Assigning a priority to the complaint based on a scale (minor scratch, single damage, multiple damages, hazard).

When querying the liquid-d1 model, the image path and user message were passed. The model would load the image, read the message and the posed questions, and then provide a response. Based on the obtained results, the probability of damage, assigned team, confidence in assignment, severity assessment, and response generation time were extracted. Since the model does not generate text responses, baseline response options were defined for each department.

Testing and Benchmark Results

Running the main script main.py showed that the model analyzed five images in an average of 100 milliseconds each.

• Plates: Classified under the fragile replacement department with a severity score of 2.6 out of 3 and a damage probability of approximately 100%.
• Cup and Saucer: Similarly classified under the fragile replacement department with a damage probability of approximately 100%.
• Cooktop: Classified under the hazardous appliances department with an appropriate message for replacement and refund.
• Luggage: Correctly classified under the luggage logistics team with a message for priority claim processing.

Thus, the model provided solutions and severity assessments for all issues in under 0.5 seconds, processing visual and textual data simultaneously and generating zero output tokens on average in 0.1 seconds.

In official benchmarks, the D1-3B model performed highly in tool-selection tasks, outperforming systems four times its size. The 600M version efficiently analyzes text sentiment, leading in the online discussion moderation category and surpassing models eight times its weight.

Deployment Options

Various options are available for local deployment:

• 8-bit version: 3 GB download size, optimized for Llama CPP and OAMA environments.
• 4-bit quantization: Approximately 2 GB in size, suitable for mobile devices.
• Uncompressed base models: Require around 8 GB of VRAM for maximum performance.

Advantages and Limitations

The main advantage of the model is the absence of processing delays, which allows it to replace larger models (7-70B) in routing tasks. However, it should be noted that D1 is exclusively a classification engine and cannot generate code or compose emails. The model is designed to obtain structured probability metrics.

Conclusion

Liquid D1 is an effective solution for instant structured routing and security filtering without the high computational costs of traditional text models. The model can act as the first layer of defense in an application architecture, analyzing incoming requests, checking them against security rules, and activating the main processing engine only when necessary, which significantly reduces computational costs.

Code files and documentation are available for setting up and running the project via the first link in the description and pinned comments.