**Breaking News: Liquid AI Disrupts AI Landscape with LFM2.5-1.2B-Considering – A 1.2B-Parameter Model That Fits Under 1 GB on Device**
In a game-changing move, Liquid AI has unveiled LFM2.5-1.2B-Considering, a 1.2 billion parameter reasoning model that can run entirely on-device while taking up less than 1 GB of storage space. This breakthrough means that complex reasoning processes can be performed on your own device, no longer requiring massive infrastructure or data centers.
**What’s so special about LFM2.5-1.2B-Considering?**
This model is part of the LFM2.5 family of Liquid Foundation Models, which builds upon the earlier LFM2 structure with more pre-training and multi-stage reinforcement learning for edge deployment. LFM2.5-1.2B-Considering is specifically designed for text-only and broad-coverage applications, making it perfect for tasks such as:
* Agentic workflows
* Information extraction
* Retrieval-augmented generation
**Under the Hood: Key Features and Benchmarks**
Here are some key stats to get you excited:
* 1.17B parameters (reporting as a 1.2B-class model)
* 16 layers with 10 double-gated LIV convolution blocks and 6 GQA blocks
* Coaching budget of 28T tokens
* Context size of 32,768 tokens
* Vocabulary dimension of 65,536
* Supports 8 languages: English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish
LFM2.5-1.2B-Considering has shown strong performance on various benchmarks, with scores of:
* 87.96 on MATH 500
* 85.60 on GSM8K
* Competitive performance with Qwen3-1.7B in thinking mode with fewer parameters
**The Secret to Avoiding Doom Loops**
We’ve all been there – stuck in an infinite loop of repetitive outputs. Liquid AI has tackled this issue with a multi-stage training pipeline, including:
* Mid-training with reasoning traces
* Supervised fine-tuning
* Desire alignment with 5 sampled candidates and 1 grasping candidate
* RLVR with n-gram penalties
This approach has reduced doom loops from 15.74% to a mere 0.36%.
**Inference Performance and Hardware Footprint**
Good news for those with limited resources – LFM2.5-1.2B-Considering can decode at an impressive:
* 239 tokens per second on an AMD CPU
* 82 tokens per second on a mobile NPU
* All while running under 1 GB of memory
**What Does This Mean for You?**
LFM2.5-1.2B-Considering is a major step forward in AI development, and its impact will be felt across industries such as:
* Natural Language Processing
* Information Extraction
* Retrieval-augmented generation
You can access or host the model via various providers and platforms, including cloud & API providers and model repositories (self-hosting). The weights are also available in various formats for those who want to run the model locally or on their own infrastructure.
Read the original article here to stay up-to-date on this groundbreaking development:
Stay ahead of the curve with Liquid AI’s LFM2.5-1.2B-Considering – a 1.2 billion parameter model that’s breaking new ground in AI capabilities and usability.
