---

license: apache-2.0

language:

- multilingual

base_model:

- LiquidAI/LFM2.5-1.2B-Instruct

tags:

- large-language-model

- instruction-tuned

- multilingual

- reasoning

- text-generation

- efficient-model

- gguf

---

LFM2.5-1.2B-Instruct

LFM2.5-1.2B-Instruct is a compact instruction-tuned language model developed by Liquid AI, designed to provide efficient natural language understanding, instruction following, reasoning, and text generation while maintaining a lightweight deployment footprint. This repository contains GGUF quantized variants of the model optimized for efficient local inference using llama.cpp.

Unlike multimodal or domain-specific models, LFM2.5-1.2B-Instruct is a text-only large language model built for general-purpose language tasks. Its compact architecture enables practical deployment on consumer hardware while delivering strong conversational ability, structured generation, multilingual understanding, and instruction-following performance.

The quantized formats significantly reduce memory requirements while preserving language modeling quality, making the model well suited for local assistants, workflow automation, embedded AI applications, and resource-efficient inference.

---

Model Overview

• Model Name: LFM2.5-1.2B-Instruct
• Base Model: LiquidAI/LFM2.5-1.2B-Instruct
• Architecture: Decoder-Only Transformer
• Parameter Count: 2 Billion Parameters
• Modalities: Text
• Primary Languages: Multilingual
• Developer: Liquid AI
• License: Apache 2.0

---

Quantization Formats

This repository provides various GGUF quantized versions of the LFM2.5-1.2B-Instruct model optimized for efficient local inference using llama.cpp.

IQ3_M

• Size reduction of approx 75.76% (540.54 MB) compared to 16-bit (2.18 GB)
• Compact 3-bit quantization optimized for ultra-low-memory language model inference
• Suitable for lightweight conversational AI, embedded deployments, and resource-constrained environments
• Enables efficient execution of instruction-following and text-generation workloads on consumer hardware
• Complex reasoning and long-form generation quality may be moderately reduced compared to higher-precision variants

IQ4_NL

• Size reduction of approx 70.08% (667.52 MB) compared to 16-bit (2.18 GB)
• Optimized 4-bit non-linear quantization balancing language quality with memory efficiency
• Recommended for production conversational AI, reasoning, summarization, and structured text-generation tasks
• Preserves instruction-following capability and contextual consistency across diverse NLP workloads
• May require slightly increased computational resources during inference

IQ4_XS

• Size reduction of approx 71.37% (637.65 MB) compared to 16-bit (2.18 GB)
• Balanced 4-bit quantization delivering an effective compromise between inference efficiency and response quality
• Suitable for local AI assistants, workflow automation, question answering, and multilingual text generation
• Provides dependable performance across a broad range of practical language understanding tasks
• Recommended for production deployments requiring efficient inference with stable output quality

Q6_K

• Size reduction of approx 58.84% (918.24 MB) compared to 16-bit (2.18 GB)
• Higher-precision 6-bit K-Quant format optimized for preserving reasoning capability and language generation fidelity
• Better suited for analytical reasoning, coding assistance, structured generation, and complex conversational workflows
• Retains more of the original model's language understanding capability compared to lower-bit variants
• Recommended when response quality is prioritized over maximum memory savings

---

Training Background (Original Model)

LFM2.5-1.2B-Instruct is trained with an emphasis on efficient language modeling, multilingual understanding, instruction following, and conversational reasoning across diverse text corpora.

Pretraining

• Large-scale language pretraining using multilingual text datasets spanning diverse domains and writing styles
• Focus on contextual language understanding, knowledge acquisition, and robust text representation learning
• Optimized for downstream conversational AI, reasoning, summarization, and text-generation tasks

Instruction Tuning

• Further refined using instruction-following and dialogue-oriented datasets
• Enhanced for conversational consistency, structured response generation, and reasoning tasks
• Improved performance across question answering, summarization, workflow automation, and general assistant applications

---

Key Capabilities

• Instruction Following

Accurately follows natural language instructions across a broad range of tasks.

• Conversational AI

Generates coherent and context-aware responses for interactive dialogue.

• Reasoning

Supports logical reasoning and multi-step problem solving.

• Multilingual Understanding

Understands and generates text across multiple languages.

• Structured Text Generation

Produces well-organized outputs suitable for automation and downstream processing.

• Efficient Local Deployment

Quantized variants enable practical inference on consumer hardware.

---

Usage Example

Using llama.cpp

./llama-cli \
  -m SandLogicTechnologies/LFM2.5-1.2B-Instruct_IQ4_NL.gguf \
  -p "Summarize the following technical document and list the key takeaways."

---

Recommended Usecases

• Conversational AI

Build lightweight local AI assistants.

• Question Answering

Answer factual and instructional queries across diverse topics.

• Content Summarization

Generate concise summaries of long-form documents and articles.

• Workflow Automation

Produce structured text outputs for enterprise automation pipelines.

• Educational Applications

Support tutoring, explanations, and interactive learning experiences.

• Research & Experimentation

Evaluate efficient language models for local inference and edge AI deployments.

---

Acknowledgments

These quantized models are based on the original work by the Liquid AI development team.

Special thanks to:

• The Liquid AI team for developing and releasing the LFM2.5-1.2B-Instruct model.
• Georgi Gerganov and the llama.cpp open-source community for enabling efficient quantization and inference through the GGUF format.

---

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/