Liquid AI: A Compact Model Challenges Google and Triumphs
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
The world of artificial intelligence has recently been shaken by an unexpected performance: a model from Liquid AI, the LFM2.5–8B-A1B, has outperformed Google's much larger Gemma 4 26B model. This Liquid AI model, which activates only 1.5 billion parameters, was able to run on a simple laptop while achieving an impressive score of 88.07 on the Tau²-Telecom benchmark. In comparison, Google's model, with its 26 billion parameters, only reached 42.11, a gap of 46 points.
The Innovative Architecture of Liquid AI
The Liquid AI model is based on an 8 billion parameter MoE (Mixture of Experts) architecture, optimized for multi-step tool usage. This approach is particularly suited for privacy-sensitive tasks and requires reliable abstention, rather than focusing solely on general rankings. The LFM2.5–8B-A1B architecture stands out by using short-range convolutional layers with gates, preferred over attention mechanisms, which enhances the reasoning of tools and agents without burdening inference costs on peripheral hardware.
Compelling Results
Benchmarks show that the LFM2.5–8B-A1B model offers significant gains over the previous generation LFM2. It also outperforms the Gemma 4 26B on the Tau²-Telecom benchmark, notably by reducing hallucinations and uncertainties. These results highlight the effectiveness of the Liquid AI model in contexts where accuracy and reliability are crucial.
Deployment and Limitations
In terms of deployment, the model utilizes technologies such as vLLM, Transformers, and llama.cpp/quantized GGUF on CPU. While it excels in dispatching agentic tools, it shows weaknesses in programming and tasks requiring a vast knowledge base. The article advises prioritizing this model for private tool calls on-device, while opting for other solutions for broader coding or information retrieval needs.
In conclusion, Liquid AI has designed a model that acts like a "scalpel" for agents, offering a specialized and efficient alternative to the "Swiss army knives" of more generalist models. This model was released on May 28, 2026, marking a turning point in the efficiency of AI models on less powerful devices.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.