On October 27, Liquid AI introduced LFM3, the latest generation of its Liquid Foundation Models, bringing significant advancements in model architecture, instruction following, reasoning, and multimodal understanding. The release introduces six models built around a shared foundation, spanning text, vision, audio, and specialized decision-making, with each designed to deliver stronger capabilities within a relatively small parameter footprint.

Today, we're bringing two of those models to Oscilla: LFM3-1.7B and LFM3-VL-2B, now available on iPhone and Mac. Both build on Liquid AI's previous generations while introducing meaningful improvements to how smaller models process information, follow complex instructions, and interpret visual content. For users exploring compact AI models, LFM3 offers an interesting combination of efficiency and capability that extends beyond a typical incremental model update.

What makes this release particularly notable is the progress Liquid AI has achieved without increasing model size. LFM3-1.7B delivers stronger results than its 2.6-billion-parameter predecessor across a range of benchmarks, while LFM3-VL-2B introduces substantial improvements in visual reasoning with fewer parameters than the previous vision model. Together, they represent an important step forward for models designed to operate efficiently on personal hardware.

A New Architecture Behind LFM3

At the center of LFM3 is a redesigned architecture that builds on the hybrid approach introduced with LFM2. Instead of relying entirely on traditional transformer attention, LFM3 combines three different mechanisms: gated short convolutions, sliding-window attention, and global attention. This approach, referred to as a tri-hybrid architecture, allows the model to process information efficiently while preserving the ability to interpret relationships across longer sequences.

The 1.7-billion-parameter text model contains 24 layers, consisting of 13 gated convolution layers, five sliding-window attention layers, and six global attention layers. The convolution layers support efficient processing of sequential information, sliding-window attention focuses on nearby context, and global attention provides a mechanism for connecting information across a broader input. Liquid AI also introduced refinements to attention gating, normalization, and positional encoding, intended to improve model quality without introducing unnecessary computational overhead.

Figure 1. Inside LFM3's tri-hybrid architecture

Layer composition of LFM3-1.7B. The diagram illustrates proportions, not layer execution order. Source: Liquid AI.

The training process has evolved as well. LFM3's base model was pretrained on 22 trillion tokens, with additional emphasis on scientific papers, mathematical problems, software repositories, and structured examples of tool use. The final 3.5 trillion tokens were part of a higher-quality training stage that placed greater emphasis on agentic interactions and diverse tool calls. Liquid AI also introduced a more extensive post-training process involving supervised fine-tuning, reinforcement learning, and distillation from specialized models trained for different tasks.

The result is a model family that places greater emphasis on following instructions, working with structured information, and performing multi-step tasks. According to Liquid AI, LFM3-1.7B outperforms LFM2.5-2.6B across numerous evaluations despite using approximately 35% fewer parameters. For compact models, where memory and processing resources remain important considerations, achieving stronger performance with a smaller architecture is one of the most significant developments in this release.

LFM3-1.7B: More Capability in a Smaller Model

The first model joining Oscilla is LFM3-1.7B, a text-based language model developed for instruction following, reasoning, coding, and agentic applications. With approximately 1.7 billion parameters, it occupies a relatively small place in the broader language model landscape, but its performance improvements demonstrate how quickly the capabilities of compact models are evolving.

One of the most notable areas of improvement is instruction following. On IFBench, a benchmark designed to evaluate a model's ability to follow detailed instructions, LFM3-1.7B achieved a score of 73.03 compared with 60.48 for LFM2.5-2.6B. On BFCL V4, which measures function-calling capabilities, the new model scored 71.39 compared with 55.89 for its predecessor. Liquid AI also reported meaningful improvements in coding and agentic evaluations, including LiveCodeBench v6 and LiveClaw.

Figure 2. LFM3-1.7B outperforms its predecessor across key evaluations

Comparison using IFBench, BFCL V4, LiveCodeBench v6, and LiveClaw, respectively. Higher is better. Different benchmarks measure different capabilities and should not be interpreted as a single performance scale. Source: Liquid AI's published evaluations

These results are particularly relevant because successful AI interactions often depend on more than producing a fluent answer. A model needs to interpret the user's intent, respect constraints, follow formatting requirements, and maintain consistency across multiple instructions. Improvements in these areas can make smaller models more practical for tasks involving structured outputs, programming assistance, and workflows that require more deliberate execution.

From Instruction Following to Agentic Workflows

Liquid AI has placed particular emphasis on agentic capabilities with LFM3-1.7B. Its post-training process includes examples of tool use and multi-step tasks, with the goal of making the model better suited for applications that interact with external tools and software environments. In Liquid AI's published evaluations, the model achieved particularly strong results in function calling and agent-oriented benchmarks, suggesting that compact models are becoming increasingly viable for tasks previously associated with substantially larger systems.

This distinction is important because a model's ability to generate text and its ability to operate reliably within an agent framework are not necessarily the same. Agentic applications require models to select appropriate actions, produce structured arguments, interpret tool results, and determine what to do next. While LFM3-1.7B's training provides a foundation for these capabilities, the functionality ultimately depends on the application and tools surrounding the model.

Within Oscilla, LFM3-1.7B is available for conversations, writing, summarization, coding assistance, and other text-based interactions. Its improvements in instruction following and structured reasoning make it an interesting addition to the model library, while the broader agentic capabilities described by Liquid AI provide an area for continued exploration as Oscilla evolves.

Longer Context Without a Larger Model

LFM3-1.7B also introduces a native context window of up to 256,000 tokens, an unusually large capacity for a model of this size. A larger context window allows more information to be included within a single interaction, whether that involves lengthy documents, extended conversations, or complex instructions containing multiple pieces of reference material.

However, architectural support for a large context window does not mean every device can use that entire capacity efficiently. Context length affects memory consumption and processing requirements, particularly during extended conversations. Actual limits within Oscilla depend on the hardware, available memory, model configuration, and runtime.

Liquid AI reports that the text model's weights occupy approximately 1 GB at 4-bit precision, providing a relatively compact starting point for deployment. In its published llama.cpp measurements, LFM3-1.7B also delivered decoding speeds between 1.8 and 2.8 times those of the next-fastest models in its selected comparisons. While those results should not be treated as direct measurements of Oscilla's MLX implementation, they highlight the efficiency goals behind the new architecture.

LFM3-VL-2B: Advancing Visual Understanding

Alongside the text model, we're introducing LFM3-VL-2B, Liquid AI's latest compact vision-language model. Built on the same architectural foundation as LFM3-1.7B, the vision model extends language understanding to visual information, allowing users to interact with images through natural-language instructions.

Vision-language models face a different challenge from conventional text models. Beyond processing written instructions, they must identify visual details, interpret text within images, recognize spatial relationships, and connect those observations to the question being asked. Reading a chart, for example, requires more than identifying labels or extracting numbers. The model needs to understand how the information is organized and reason about the relationships between different elements.

LFM3-VL-2B introduces a new vision encoder developed by Liquid AI that combines convolution and attention mechanisms to process visual information more efficiently. According to the company, the encoder requires approximately one-third of the computation used by the SigLIP2 encoder in previous Liquid vision models. This reduction is particularly relevant to compact vision-language systems, where image processing can represent a substantial portion of the computational workload.

The performance improvements extend across a range of visual tasks. In Liquid AI's published evaluations covering 62 vision benchmarks, LFM3-VL-2B achieved an overall average of 71.4 compared with 57.6 for LFM2.5-VL-3B. The new model also uses approximately 1.9 billion parameters, compared with 3.1 billion in the earlier model, demonstrating improvements in both model size and evaluated capability.

Stronger Results Across Documents, Charts, and Visual Reasoning

Some of the largest improvements appear in areas that involve interpreting complex visual information. On Liquid AI's document, OCR, and chart evaluation category, LFM3-VL-2B achieved a score of 83.9 compared with 71.2 for LFM2.5-VL-3B. Its mathematical and logical reasoning score increased from 40.6 to 68.8, while GUI understanding improved from 66.0 to 83.1.

Figure 3. A significant improvement in visual understanding

Liquid AI's published benchmark category averages for documents and charts, mathematical and logical reasoning, GUI understanding, and visual grounding. Higher is better. These scores were not measured inside Oscilla.

These advances are particularly interesting because they represent tasks where visual recognition must be combined with reasoning. Understanding a screenshot involves recognizing interface elements and how they relate to one another. Interpreting a document can require identifying relevant information across different regions of a page. Working with a graph or diagram requires the model to understand not only what is present but also how the information should be interpreted.

Within Oscilla, users can explore LFM3-VL-2B by attaching an image to a conversation and asking questions about its contents. Whether examining a screenshot, interpreting a chart, or exploring information contained in a photograph, the model introduces another option for working with visual content directly through the app.

Liquid AI has also expanded the underlying model's capabilities to include video understanding, explicit reasoning, and more advanced spatial interpretation. Those capabilities are part of the broader model architecture, although Oscilla's current integration focuses on image-based interactions rather than the full range of supported video and multimodal features.

Choosing Between LFM3-1.7B and LFM3-VL-2B

Although the two models share the same underlying LFM3 foundation, they are designed for different types of interactions. LFM3-1.7B focuses on text-based tasks, offering improvements in instruction following, coding, and structured reasoning. LFM3-VL-2B extends that foundation with the ability to understand images and combine visual observations with natural-language instructions.

Figure 4. Comparing the LFM3 models available in Oscilla

Specifications from Liquid AI. Supported inference configurations and available context within Oscilla depend on the device and runtime.

For users working primarily with writing, questions, summaries, or text-based tasks, LFM3-1.7B offers a focused and efficient option. Users who regularly work with screenshots, photographs, diagrams, or other visual material may find LFM3-VL-2B more suitable. The additional vision capabilities introduce a different set of possibilities, although they also bring different processing and memory requirements depending on the task.

Neither model is intended to replace every other option available in Oscilla. Larger models continue to offer advantages in certain areas, particularly more demanding knowledge and reasoning tasks, while different compact architectures may perform better depending on the type of work being performed. What makes LFM3 interesting is the improvement in the balance between size, efficiency, and capability, especially when compared with previous generations from Liquid AI.

What's Next for LFM3 in Oscilla?

The initial LFM3 release includes a broader collection of models beyond the two now available in Oscilla. Liquid AI has also introduced LFM3-Audio-2B, LFM3-Omni-2B, and two specialized decision models. These extend the LFM3 foundation into speech, combined audio-visual understanding, and structured decision-making.

The audio and omni models are particularly interesting because they expand how compact models can interact with different forms of information. LFM3-Audio-2B introduces capabilities for speech recognition, speech generation, and audio understanding, while LFM3-Omni-2B brings text, vision, and audio together within a single multimodal system. The specialized decision models take a different approach, focusing on producing structured decisions rather than conventional conversational responses.

We're continuing to evaluate additional LFM3 models for potential inclusion in Oscilla. As with any model integration, compatibility with our inference runtime, memory requirements, performance, and the overall user experience will determine what we introduce next. Our focus remains on supporting models in a way that makes their capabilities practical and reliable rather than simply expanding the number of available downloads.

For now, LFM3-1.7B and LFM3-VL-2B are available in Oscilla on iPhone and Mac, giving users access to two distinct capabilities from Liquid AI's latest generation.

Experience LFM3 in Oscilla

LFM3 represents a meaningful advancement in the development of compact language and multimodal models. Its redesigned architecture, improved training process, and stronger benchmark results demonstrate how much capability can be achieved within a relatively small parameter footprint. The improvements in instruction following, agentic tasks, and visual understanding are particularly noteworthy, especially as these capabilities become increasingly important in everyday AI applications.

We're excited to bring LFM3-1.7B and LFM3-VL-2B to Oscilla and to continue exploring what the broader model family makes possible. Both models are available through the model library in Oscilla for iPhone and Mac, where users can download them and begin exploring their capabilities.

For more information about the LFM3 architecture, benchmarks, and complete model family, visit Liquid AI's official LFM3 announcement.