Open-weight local LLM

LFM2.5-8B-A1B

Liquid AI hybrid model built for on-device assistants. 8.3B total / 1.5B active, 128K context, tool use, GGUF, ONNX, MLX, llama.cpp and LM Studio support. Open-weight under LFM 1.0.

Laptop ready 8 GB RAM Q4_K_M On-device personal assistant
Parameters
8.3B (1.5B active)
Minimum RAM
8 GB
Model size
5.2 GB
Quantization
Q4_K_M

Can LFM2.5-8B-A1B run locally?

LFM2.5-8B-A1B is a good fit for normal laptops and compact desktops with 8 GB RAM or more.

Search for lfm2.5-8b-a1b in LM Studio or another GGUF-compatible runtime.

chatcodereasoningspeedstandardgeneral

Install path

01
Check RAM fitMinimum 8 GB RAM. Start with the Q4_K_M quant.
02
Load the modelSearch lfm2.5-8b-a1b in LM Studio.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Strengths

  • Designed specifically for on-device personal assistants and local agent workflows
  • Only 1.5B active parameters at inference despite 8.3B total parameters
  • 128K context window for long local sessions and document-heavy prompts
  • Day-one GGUF, ONNX, MLX, llama.cpp and LM Studio support
  • Strong fit for structured outputs, tool use and lightweight agentic tasks
  • Runs on mainstream 8-16 GB machines with quantized weights

Limitations

  • LFM 1.0 is a custom open-weight license, not Apache 2.0
  • Liquid AI notes it is not the best fit for heavy programming or knowledge-heavy QA without retrieval
  • Hybrid architecture may need recent runtimes for best performance
  • Still a small active-parameter model; larger 20B-30B class models can beat it on raw quality

Best use cases

  • On-device personal assistant
  • Local OpenClaw agents with tool calls
  • Structured output workflows
  • Fast multilingual chat on laptops
  • Long-context local note and document workflows
  • Apple Silicon inference through MLX

Capability profile

speed
9
quality
8
coding
8
reasoning
8

Technical notes

Developer
Liquid AI
License
LFM 1.0
Context window
128,000 tokens
Architecture
Hybrid Liquid Foundation Model with 8.3B total parameters and 1.5B active parameters. The model mixes double-gated LIV convolution layers with grouped-query attention.

Similar models to compare

Where to go next