Liquid LFM 2
Liquid LFM 2
Not a transformer, and on a phone or a laptop CPU that turns out to matter.
Updated
The read
Liquid builds on a different architecture entirely, and the payoff is decoding speed on hardware without a GPU: roughly twice what a comparable transformer manages on a CPU. For an on-device assistant, an embedded product, or anything that has to work with the network off, that is the whole ballgame. These are small models with the knowledge ceiling small models have, and the architecture is unfamiliar to most tooling.
Where it shines
- Roughly twice the CPU decoding speed of comparable transformers
- Runs genuinely offline on phone-class hardware
- Small memory footprint, so it fits inside an app
Worth knowing first
- Small models, with a correspondingly low capability ceiling
- Non-transformer architecture that most tooling does not expect
Index score
66.7
Provisional. This score is read from the product's public capability, not from the one-prompt rebuild the fully reviewed entries went through, so treat it as a placing rather than a verdict.
- Rank
- 53 of 53
- Builds
- LLMs
- Category
- Small and fast
- Output
- Small models sized for phones and edge devices
- Pricing
- Free tier
- Best for
- On-device inference with no GPU and no network
Scores are our own editorial judgement, weighted craft 55, speed 10, control 15 and value 20.
Also worth a look
Full index- 13GPT-5.6 LunaGPT-5.6 LunaProvisionalFrontier-family manners at a price that lets you stop counting tokens.Small and fastPaid only83.3
- 28Claude Haiku 4.5Claude Haiku 4.5ProvisionalAnthropic's cheap, quick one, for the work that does not need a genius.Small and fastPaid only78.9
- 31Mistral Small 4Mistral Small 4ProvisionalThe best small open model to build a product on, and the licence has no catch.Small and fastFree tier78.1
