LFM2 Models: Liquid AI for Agentic Workflows
Liquid AI's LFM2 takes a different path from the big transformer families. Its 24B-A2B flagship is a hybrid mixture-of-experts model: 24B parameters in total, with only 2B active for each token. The result is large-model knowledge with small-model speed, loading in roughly 14GB on any Mac with 16GB of unified memory. Liquid AI aims it squarely at agent work. It powers LocalCowork, an open-source agent app with 75 MCP tools. If you want a local model that reliably calls tools and follows structured workflows without touching the cloud, LFM2 deserves a look.
- developer
- Liquid AI
- local models
- 2
- size range
- 8.3B-24B
- RAM range
- 8-24 GB
All LFM2 models
2 local models| Model | Size | Quant | VRAM | Min RAM | Best For | Quality | Ollama |
|---|---|---|---|---|---|---|---|
| LFM2.5 8B-A1B | 8.3B | Q4_K_M | 5.5 GB | 8 GB | On-device agents, tool calling, multilingual chat | 84 | |
| LFM2 24B-A2B Instruct | 24B | Q4_K_M | 14 GB | 24 GB | Local AI agents, privacy-first tool calling, MCP workflows | 85 |
Key features
- Hybrid MoE design: 24B total parameters, only 2B active per token
- Loads in about 14GB, fits any 16GB Mac
- Built for agent workflows, tool calling, and structured output
- Powers LocalCowork, an open-source agent app with 75 MCP tools
- Strong CPU throughput thanks to the small active parameter count
- Designed for zero-cloud, privacy-sensitive workflows
Device compatibility
Which LFM2 models can run on each device class, based on minimum RAM requirements.
| Model | iPhone | Air | Pro | Studio | Mini | |
|---|---|---|---|---|---|---|
| LFM2.5 8B-A1B (8.3B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| LFM2 24B-A2B Instruct (24B) | No | Possible | Possible | Possible | Possible | Possible |
RAM requirements
Largest dense Q4 model that fits each memory tier
Get told when a better model fits your MacBook Air M5 16 GB
Get one email when a new open-weight model is a better fit for this machine.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Config reports
| Source | Machine | tok/s | Quant | Runtime | Context | Note |
|---|---|---|---|---|---|---|
| Speed test | Apple M2 10-core GPU 16 GBLFM2.5 8B-A1B | 58.2 | Q4_K_M | Ollama | - | 1 run |
Speed test: runs submitted through the ModelFit speed test tool, one median per machine.
Share your config
Frequently asked questions
What is the LFM2 24B-A2B MoE and what Mac can run it?
It is a sparse mixture-of-experts model: 24B total parameters with only 2B active per token. It loads in about 14GB, so it runs on any Mac with 16GB of unified memory: a 16GB MacBook Air, base MacBook Pro, or Mac Mini.
What makes LFM2 different from other models?
LFM2 uses a hybrid architecture instead of a standard dense transformer. The MoE design activates a small slice of the model per token, which keeps inference fast while retaining 24B-scale knowledge. It is tuned for tool calling and agent reliability.
Is LFM2 good for general chat?
It handles chat fine but shines on structured tasks, tool use, and agent workflows. For pure conversational quality at a similar RAM budget, Qwen or Gemma alternatives are the stronger picks.
What Ollama command runs LFM2?
Run `ollama run lfm2:24b-a2b`. You need a 16GB machine at minimum, since the model itself occupies about 14GB once loaded.
