Try Delta
← All models

LiquidAI2026

LFM 2.5

Edge & agentic · dense & MoE

LiquidAI’s LFM 2.5 line spans tiny edge models through MoE variants for agents. Built for low-latency local inference on phones, laptops, and workstations.

  • Fits fromPhone / edge · 4 GB+
  • Contextup to 33K
  • Sizes4
  • RunsLocal on Delta
EdgeAgentsReasoningEdge devicesAgentsLow latency

Sizes

LFM 2.5 1.2B

922 MB • 33k ctx • Q4_K_M

LFM 2.5 2.6B

1.8 GB • 33k ctx • Q4_K_M

LFM 2.5 8B-A1B

5.5 GB • 33k ctx • Q4_K_M

LFM 2.5 24B-A2B

14 GB • 33k ctx • Q4_K_M

Memory at context length

Approximate working set for LFM 2.5 1.2B as context grows.

4K
1.4 GB
8K
1.5 GB
16K
1.6 GB
33K
1.8 GB

Related families

Nemotron 3 Nano

NVIDIA’s compact agentic 4B: long-context assistants that stay on a laptop.

Qwen 3.8

Qwen 3.8 Distill models: highly efficient and capable models optimized for edge and desktop.

GLM

Agentic reasoning for software, browser automation, and multi-turn tools, locally.