Featured image courtesy of Liquid AI.

Liquid AI has introduced LiquidCodec, a streaming audio codec with approximately 74 million parameters. This new component, which runs at 12.5 Hz with 12 codebooks, can encode audio about 361 times faster than real time on a single CPU thread, according to reporting by MarkTechPost. The company states LiquidCodec powers its LFM3-Audio-2B and LFM3-Omni-2B models.

A Series of Swift Introductions

The release of LiquidCodec on November 12 continues a rapid sequence of product introductions from Liquid AI. Since early October, when the company first appeared on our radar, it has brought new AI capabilities to market quickly. Just weeks earlier, on October 14, Liquid AI launched the LFM2-5-24B-A2B model. Two days later, on October 16, it unveiled the D1-Ultra decision model.

This pattern of quick, successive launches shows a strategy of broad, agile development. In less than a month, Liquid AI has moved from a general-purpose model to a specialized decision-making AI, and now to a foundational audio processing component. The D1-Ultra, for instance, suggests an interest in AI systems designed for specific, outcome-oriented tasks. It moves beyond generic language generation.

Building Blocks for Broader AI Systems

LiquidCodec powers other, larger models like LFM3-Audio-2B and LFM3-Omni-2B, an approach where core, high-performance components build more complex AI systems. The ability to encode audio at 361 times real-time speed on a single CPU thread shows an emphasis on efficiency and scalability. These are critical for deploying advanced AI.

LiquidCodec’s parameter count, about 74 million, places it in a category of specialized models for specific tasks. It is not a multi-billion parameter general-purpose LLM. This focus on a smaller, highly optimized component suggests Liquid AI is building a modular AI architecture. Such an approach could enable faster iteration and easier integration of new capabilities, allowing the company to assemble and launch diverse AI products quickly.

Competitive Strategy and Future Releases

In a competitive field where major players often announce large, monolithic models, Liquid AI’s strategy of delivering varied specialized models and foundational components in rapid succession stands out. The company has developed and shipped different types of AI—from what appears to be a general model, to a decision-focused system, and now a high-performance audio codec—all in a very short timeframe. This broad scope, coupled with the component efficiency seen in LiquidCodec, could allow Liquid AI to target multiple niches or build out extensive multimodal systems from the ground up.

The swift introduction of distinct AI products, from LFM2-5-24B-A2B to D1-Ultra and now LiquidCodec, implies a strong development pipeline and a willingness to explore different AI applications simultaneously. As the AI industry continues to evolve, companies that can quickly adapt and deliver diverse, high-performance tools gain strategic advantages. Liquid AI’s early moves suggest a company aiming to establish a presence across various AI domains. It does not specialize narrowly from its inception.

The specific mention of LiquidCodec powering LFM3-Audio-2B and LFM3-Omni-2B, models not yet fully detailed, indicates Liquid AI builds a portfolio of connected AI systems. These names suggest further ventures into audio processing and potentially broad, omni-modal AI. The company appears to be laying technical groundwork. This will allow it to expand its offerings into richer, multimodal experiences, with efficient components like LiquidCodec as critical components.

Key Details

What is LiquidCodec?

LiquidCodec is a streaming audio codec released by Liquid AI, featuring approximately 74 million parameters.

How fast does LiquidCodec encode audio?

LiquidCodec encodes audio about 361 times faster than real time on a single CPU thread.

What other models has Liquid AI released recently?

Liquid AI has also released the LFM2-5-24B-A2B model on October 14 and the D1-Ultra decision model on October 16.

Compiled by Launch91 Desk from the sources linked above. More about Launch91.