Liquid AI

LFM2.5 VL 3B

Liquid AI

Liquid’s larger LFM2.5 vision model integrates a 400M SigLIP2 NaFlex encoder with the 2.6B language backbone. It strengthens screen understanding, grounding, multi-image processing, and function calling. The post-trained model answers directly without long reasoning traces, making it suitable for interactive visual assistants. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 2.6B

Liquid AI

This dense agent model scales pretraining to about 34T tokens and adds a dedicated 128K context-extension phase. Post-training combines teacher specialization, on-policy distillation, and reinforcement learning for planning and chained tool calls. The model is designed for smartphones, CPUs, and GPU serving, with downloadable base and post-trained checkpoints. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 230M

Liquid AI

Liquid’s smallest LFM2.5 text tier is designed for inexpensive local inference and downstream fine-tuning. Its training includes 19T tokens and a 32K context extension, followed by distillation, preference optimization, and reinforcement learning. It can select tools from natural-language instructions, but demonstrations of robot control rely on external low-level skills. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 8B A1B

Liquid AI

The sparse LFM2.5 tier combines expert routing, short convolutions, and attention for local agents. It extends context to 128K and expands the tokenizer for non-Latin scripts. Unlike its preceding LFM2 checkpoint, the post-trained version always generates reasoning traces. The model card’s measured parameter counts are more precise than the rounded A1B product name. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 VL 450M

Liquid AI

The compact visual tier improves multilingual image understanding, grounding, and instruction following through extended pretraining and reinforcement learning. It adds bounding-box prediction and text function calling to visual question answering. Its small footprint supports real-time edge applications that turn image observations into structured language outputs. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 350M

Liquid AI

This separately launched compact tier adds extended pretraining and reinforcement learning to Liquid’s 350M architecture. It focuses on structured output, extraction, and tool selection rather than lengthy general reasoning. Both base and post-trained weights are available for developers who need low-cost specialized inference on small edge devices. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2 24B A2B

Liquid AI

Liquid’s largest second-generation language model combines sparse experts, short convolutions, and grouped-query attention. It targets tool-using agents that can run on consumer computers while retaining cloud-serving options. The checkpoint card documents a 32K trained context; larger configuration ceilings are not treated as validated operating limits. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 1.2B Thinking

Liquid AI

The Thinking checkpoint is a distinct reasoning tier that produces intermediate thinking traces before its final answer. It targets offline mathematics, instruction following, and tool selection on memory-constrained hardware. Quantized deployment can fit below 1GB, although exact memory and throughput depend on the hardware, precision, and context. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 Audio 1.5B

Liquid AI

LFM2.5 Audio is an end-to-end conversational audio-language model with text and speech generation. It supports low-latency spoken interaction, speech recognition, and audio-conditioned answers on local hardware. This page covers its dialogue capabilities rather than treating transcription alone as a separate language model. Japanese specialization is grouped as a variant. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 VL 1.6B

Liquid AI

The first LFM2.5 vision-language checkpoint combines the compact language backbone with image processing. It supports multilingual visual questions, multiple images, and instruction following while targeting local deployment. This is a language-generating multimodal assistant, rather than a detector or standalone image encoder. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2.5 1.2B Instruct

Liquid AI

LFM2.5 extends the compact language backbone’s pretraining and scales reinforcement learning for instruction following and tool use. This page represents the 1.2B general-purpose text tier and groups its base and Japanese language variants. It serves as the language foundation for separately released vision and audio models. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM2 8B A1B

Liquid AI

LFM2 8B A1B is Liquid AI’s first on-device mixture-of-experts model in the LFM2 generation. It combines the family’s short-convolution and attention blocks with sparse expert activation, totaling about 8.3B parameters and activating approximately 1.5B per token despite the A1B product name. The release includes downloadable weights and CPU-focused inference optimization, with 32K context. It targets multilingual conversation and private edge deployment. Custom Liquid licensing permits some commercial uses but requires separate licensing for larger companies.

LFM2 Audio

Liquid AI

LFM2 Audio is Liquid AI’s end-to-end audio and text language model for responsive local conversations. It accepts and generates both speech and text within a compact 1.5B system, enabling voice chat, translation, and language-based responses rather than only transcription or speech synthesis. The release includes weights, inference code, and a reference live conversation application. Its compact design targets phones, laptops, and other resource-constrained devices. The current generation has been superseded by LFM2.5 Audio; custom weight licensing and hardware-dependent latency still apply.

Liquid Nanos

Liquid AI

Liquid Nanos is Liquid AI’s task-specialized language model release for local extraction, translation, mathematics, retrieval-grounded answers, and tool calling. The initial set includes 350M to 2.6B models, each optimized for a narrow language workflow rather than a general assistant. Named checkpoints include English/Japanese translation, mathematical reasoning, RAG, and tool-use models. Downloadable weights and device integrations support private deployment. Reported frontier-model comparisons apply to individual specialized tasks, not overall general intelligence or arbitrary agent workloads.

LFM2 VL

Liquid AI

LFM2 VL extends Liquid AI’s efficient language architecture to image understanding. The original 450M and 1.6B models connect SigLIP2 NaFlex vision encoders to LFM2 backbones through a compact projector, supporting variable image shapes and document or scene questions. A larger 3B checkpoint followed on October 22. This page groups routine sizes while preserving the later checkpoint’s own chronology. Downloadable weights use Liquid’s custom license; local inference supports privacy and edge deployment. Image resolution and patch counts affect speed, memory, and visual detail.

LFM2

Liquid AI

The second Liquid Foundation Model generation uses gated short convolutions alongside attention to improve local inference. Its language checkpoints target instruction following, structured generation, and efficient execution on phones and personal computers. This family page groups the original dense sizes and their instruction checkpoints; later sparse tiers and multimodal models have separate pages. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

LFM 7B

Liquid AI

LFM 7B is Liquid AI’s language model release for efficient enterprise and edge intelligence. The separately released 7B model targets private enterprise chat, multilingual instruction following, coding, and agent workflows. Its architecture prioritizes efficient inference and low memory use. The original hosted generation is now deprecated in favor of newer LFMs. This page records the original language product, rather than assuming the architecture’s theoretical ability to model many signals means native audio or video support in this checkpoint. Deployment and accuracy depend on the selected model and hardware.

LFM 1

Liquid AI

LFM 1 is Liquid AI’s language model release for efficient enterprise and edge intelligence. The first-generation family includes compact dense models and a larger mixture of experts, designed around computational units inspired by dynamical systems and signal processing. Public playground access accompanied the release; open weights are not established. This page records the original language product, rather than assuming the architecture’s theoretical ability to model many signals means native audio or video support in this checkpoint. Deployment and accuracy depend on the selected model and hardware.