Liquid: LFM2-2.6B

LFM2-2.6B is Liquid AI's largest dense model, featuring dynamic hybrid reasoning and strong performance across instruction following, math, and multilingual tasks.

LFM2 sets a new standard in terms of quality, speed, and memory efficiency. — Liquid AI

Overview

Released September 23, 2025, LFM2-2.6B is a dense model (all parameters active) that outperforms similarly-sized models like Llama-3.2-3B-Instruct and SmolLM3-3B across multiple benchmarks. It's the only LFM2 model in the dense family with dynamic hybrid reasoning capabilities.

Benchmark Performance

LFM2-2.6B outperforms similar-sized models across knowledge, instruction following, and math:

Benchmark	LFM2-2.6B	Llama-3.2-3B	SmolLM3-3B	gemma-3-4b-it
MMLU	64.42	60.35	59.84	58.35
IFEval	79.56	71.43	72.44	76.85
GSM8K	82.41	75.21	81.12	89.92
MGSM	74.32	61.68	68.72	87.28
MMMLU	55.39	47.92	50.02	50.14

Recommended Use Cases

• Agentic tasks and function calling
• Data extraction
• RAG pipelines
• Creative writing
• Multi-turn conversations

Not recommended for: Knowledge-intensive tasks or complex programming (fine-tune for these)

When to Use LFM2-2.6B

Choose LFM2-2.6B when you need:

• Dense model simplicity (no MoE routing complexity)
• Dynamic hybrid reasoning for complex prompts
• Broad device compatibility
• Strong multilingual performance

Choose LFM2-8B-A1B when you need:

• Better code and knowledge capabilities
• Faster inference per quality level (MoE efficiency)

Choose LFM2-1.2B or smaller when you need:

• Minimum memory footprint
• Deployment on constrained devices

Hardware Requirements

Quantization	RAM/VRAM Required
Q4_K_M	~2GB
8-bit	~3GB
BF16	~5.5GB

Runs efficiently on smartphones, tablets, and laptops with CPU, GPU, or NPU.

Role in Series

LFM2 dense text models:

1. LFM2-2.6B: Largest dense, dynamic reasoning (this model)
2. LFM2-1.2B: Mid-size dense
3. LFM2-700M: Compact
4. LFM2-350M: Smallest

Links

• HuggingFace
• Blog Post
• Documentation
• LFM2 Technical Report