NovachronoAI/Nova-LFM-1.2B-Thinking
NovachronoAI's Nova-LFM-1.2B-Thinking is a 1.2 billion parameter language model built on the Liquid LFM-2.5 architecture, specifically engineered for efficient System 2 reasoning. It utilizes a novel Hybrid Self-Correction fine-tuning method, allowing it to pause, verify, and correct its own logic, bringing complex reasoning capabilities typically found in much larger models to edge devices. This model excels at mathematical and logical problem-solving while maintaining general knowledge, making it suitable for applications requiring robust reasoning on resource-constrained hardware.
Loading preview...
Nova-LFM 1.2B Thinking: System 2 Reasoning at the Edge
Nova-LFM 1.2B is a highly efficient "Thinking" model developed by NovachronoAI, designed to bring advanced reasoning capabilities to resource-constrained environments. Built upon the Liquid LFM-2.5 architecture, this 1.2 billion parameter model employs a unique Hybrid Self-Correction fine-tuning method, enabling it to "think" by pausing, verifying its logic, and correcting errors before generating a final response.
Key Capabilities
- System 2 Thinking: Effectively breaks down and solves multi-step mathematical and logical problems, moving beyond simple pattern matching.
- Edge-Native Performance: Optimized to run on devices with limited resources, requiring less than 3GB VRAM (FP16) or under 1GB (quantized), making it suitable for Raspberry Pis, older smartphones, and laptops.
- Balanced Performance: Achieves strong mathematical reasoning (53.5% on GSM8K) without sacrificing general knowledge (50.1% on MMLU), offering a superior balance compared to other models in its class.
Benchmark Highlights
Nova-LFM 1.2B demonstrates competitive performance, outperforming models like Llama 3.2 1B in mathematical reasoning and maintaining higher general knowledge than specialized "math-only" models. For instance, it achieves 53.5% on GSM8K and 50.1% on MMLU, contrasting with models like DeepSeek R1 Distill (1.5B) which scores 69.9% on GSM8K but suffers from significant knowledge degradation (39.2% on MMLU).
Methodology: Hybrid Self-Correction
The model's ability to self-correct stems from a fine-tuning process using a hybrid dataset that combines 11k standard Chain-of-Thought (CoT) examples with 4k synthetic self-correction traces. This teaches the model to explicitly doubt and correct its own reasoning, significantly reducing logic errors in complex tasks.