Liquid AI

An update to the small MoE from Liquid AI, increasing capabilities across the board. The small 230M version also got an update.

An update to the small MoE from Liquid AI, increasing capabilities across the board. The small 230M version also got an update.

A tiny vision model built on top of LFM2.5 and SigLIP2.

With 28T tokens for 350M parameters, this model might be the most overtrained model out there.

A relatively "big" model by Liquid AI, given their focus on edge deployments.

Liquid continued pretraining from 10T (of their 2.0 series) to 28T tokens and it shows! This model update really surprised us: In our vibe testing, it came very close to Qwen3 4B 2507 Instruct, which we use every day. And this model is over 3 times smaller! In a direct comparison against the (still bigger) Qwen3 1.6B, we preferred LFM2.5 basically every time. And this time, they released all the other variants at once, i.e., a Japanese version, a vision and an audio model.

Liquid continued pretraining from 10T (of their 2.0 series) to 28T tokens and it shows! This model update really surprised us: In our vibe testing, it came very close to Qwen3 4B 2507 Instruct, which we use every day. And this model is over 3 times smaller! In a direct comparison against the (still bigger) Qwen3 1.6B, we preferred LFM2.5 basically every time. And this time, they released all the other variants at once, i.e., a Japanese version, a vision and an audio model.

Liquid continued pretraining from 10T (of their 2.0 series) to 28T tokens and it shows! This model update really surprised us: In our vibe testing, it came very close to Qwen3 4B 2507 Instruct, which we use every day. And this model is over 3 times smaller! In a direct comparison against the (still bigger) Qwen3 1.6B, we preferred LFM2.5 basically every time. And this time, they released all the other variants at once, i.e., a Japanese version, a vision and an audio model.

A multilingual, multi-vector embedding model covering 9 languages.

A vision model by Liquid AI, who continue to churn out open models (under a restrictive license).

A small MoE by LiquidAI.

LiquidAI has also started to increase their model sizes. This model is fast but not as capable as Qwen3 in our vibe testing. It also comes with a restrictive license.

LiquidAIs first end-to-end audio model.

A vision version of LFM2.

A series of hybrid models aimed at edge devices.