Liquid AI has introduced a hybrid engine for its LFM 2.5 model, which activates only 2 billion parameters for each computational task. This enables high processing speeds with low computational costs, something that previously required expensive hardware. This article examines the key features of the model and its testing results across six complex use cases.

Architecture and Performance

The LFM 2.5 model has a total of 24 billion parameters, however, only a portion of them is utilized during operation. Even with 2 billion parameters activated, the model demonstrates impressive results: achieving a score of 48 points on comprehensive agent benchmarks, which significantly surpasses the Q1 version with 35 billion parameters.

When using an Apple M5 processor, token generation speed reaches 184 tokens per second. Average hardware configurations can support a generation speed of 64 tokens per second. The model possesses a large context window, accommodating over 100,000 tokens, and the total volume of data on which the model was trained is 36 trillion tokens.

Model Testing

Testing was conducted on the online platform playground.lquid.ai, which allows for model usage without the need for local installation. Six different use case scenarios were proposed and run in parallel.

Test 1: Inventory Analysis and Visualization

The first test evaluated the model's ability to analyze data and create visualizations. The model was provided with a list of warehouse items and tasked with calculating the quantity of each item and its percentage of the total inventory. It was then required to create a visual histogram using text characters to represent the quantity of each item. Finally, it needed to generate a heatmap in the form of a 10x85 text grid, where each inventory type is represented by a unique character, and the grid is filled sequentially to visually represent the occupied space. The model successfully completed the task, providing a table with item counts and percentages, a horizontal histogram, and a heatmap with unique characters for each item. A detailed explanation of the object distribution across the grid was also provided.

Test 2: HTML Dashboard Creation

The second test involved generating a complete HTML dashboard with clean CSS styling, without using external frameworks like React or Vue. The dashboard was to include a sidebar on the left and a main content area. The application was intended for project tracking, with the ability to input a project name and remaining hours. Upon clicking the "Submit" button, the data should instantly update in a dynamic HTML table. The model successfully created a dynamic HTML dashboard that updated upon data entry.

Test 3: Python Implementation of "Game of Life"

The third test required the implementation of "Game of Life" in Python with terminal output. It was necessary to create the classic "glider" pattern and the "blinker" oscillator in a 20x40 grid, with a 0.2-second pause between generations. Live cells were to be displayed as square blocks, and dead cells as dots. Objects exiting the grid boundaries should automatically wrap around to the opposite side. The model generated a Python script that correctly displayed "Game of Life" in the terminal with proper boundary handling.

Test 4: Kanban Board Creation

The fourth test involved creating a Kanban board for task tracking with "todo," "in progress," and "done" statuses. It required the ability to dynamically move cards between columns using a "move right" button. All elements, including HTML, CSS, and JavaScript, were to be combined into a single file without external dependencies. The model successfully implemented the Kanban board functionality, allowing tasks to be added and moved between statuses within a single HTML file.

Test 5: Vector Diagram Creation

The fifth test required the creation of a vector diagram for a project status tracking application. The diagram was to include a full arc from 0 to 100 with three color zones (red, yellow, green) and an arrow pointing to the current value of 78%. Text labels for 0, 100, and 78% were also required. The model created the diagram with the specified zones and arrow, although the arc shape could have been more semicircular.

Test 6: Terminal Sinusoid Generation

The sixth test involved creating a Python script to generate a sinusoid in the terminal using only asterisk characters. It required displaying an animation of a wave moving vertically, with a specified amplitude, cycle time, and the use of the asterisk character to draw the curve. The model successfully generated a sinusoid in the terminal using only asterisks and correctly displayed two full cycles.

Engine Architecture

The engine's input layer supports large context windows and features a vocabulary size of 128,000 words, enabling the model to comprehend diverse languages. Internal processing utilizes dual short-path convolutional layers, replacing standard large transformer blocks, which reduces memory consumption by 75%. A routing system manages 64 expert models, selecting the top four for a specific task, ensuring only 2 billion parameters are activated. The engine employs a direct "chain-of-thought" process for precise tool execution and result retrieval. The temperature is set low (0.1), and the top-k parameter limits choices to 50 candidates, ensuring logical and stable operation.

Benchmarks and Comparison

Benchmarks show that the model achieves 28% on lab automation tests, which is 3.22 times higher than Q1 35B. On agentic behavior tests, the model scores 48 points, outperforming Q1 35B, Google Jama 4, and Nvidia Neotron by 26%. The instruction following score is 97%. In function calling tests, the model scores over 80%, surpassing competitors.

However, in the areas of scientific reasoning, mathematical proofs, and code correction, Q1 35B slightly outperforms LFM 2.5. Nevertheless, on most other benchmarks, LFM 2.5 demonstrates superiority, activating only 2 billion parameters.

Local Installation and Requirements

The model is available for local installation. The 4-bit quantized version is 14 GB in size and requires 12-16 GB of VRAM. The partial offload to RAM version requires 6 GB of VRAM and 16 GB of RAM. For mobile devices, 2-bit quantized versions of approximately 10 GB are available, requiring 8-12 GB of VRAM. Local installation is done using the llama CPP command, specifying the GGF file and configuration, including a repeat penalty of 1.05 and flash attention for large contexts. 4-bit quantization preserves the accuracy of the original model.

Advantages and Disadvantages

Advantages:

• High processing speed: up to 186 tokens/sec on RTX3090 and 64 tokens/sec on CPU.
• Lack of excessive complexity: the model does not require external dependencies for creating web dashboards.
• Transparency: all benchmarks are published online.

Disadvantages:

• Temperature sensitivity: values above 0.5 can cause hallucinations. The temperature should be strictly between 0.1 and 0.2.
• Limited creativity: the model is focused on analytical and logical tasks, being inferior in creative writing and storytelling to models like Jamma 26B.
• License: the model is distributed under the proprietary Liquid AI Openweight license, not under standard open licenses.

The LFM 2.5 A2B model from Liquid AI offers powerful capabilities for working with agents and complex reasoning on budget hardware. For more information, including specifications, hardware requirements, and prompts used, please refer to the documentation at the link in the description.