A new algorithm from BitEnergy AI points to a simpler way to run some of the most demanding operations inside artificial intelligence systems. Called Linear-complexity multiplication, or L-Mul, the method is designed to swap complex floating-point multiplications for simpler integer additions.
The result, according to the study Addition is All You Need for Energy-Efficient Language Models, could be a major cut in energy consumption for specific AI model operations. The researchers say L-Mul could reduce energy use by up to 95% for element-wise floating-point tensor multiplications and by 80% for dot products.
What L-Mul Changes Inside AI Computation
Modern AI models rely on many mathematical operations to process data, compare patterns, and produce outputs. The source study focuses on replacing some floating-point multiplication work with integer addition, a simpler operation for the system to handle.
That distinction matters because the energy cost of AI is not only about model size or training time. It is also shaped by the basic operations repeated across language, vision, and reasoning tasks. If a frequent operation can be made less energy-intensive, the savings can become meaningful at scale.
L-Mul is not presented as a new model family. It is an algorithmic method that changes how certain computations are performed. In plain terms, it aims to keep AI systems doing useful work while reducing the cost of part of the math underneath.
Where The Energy Savings Could Come From
The study reports two main energy-reduction figures. For element-wise floating-point tensor multiplications, L-Mul could cut energy use by up to 95%. For dot products, the reported reduction is 80%.
Those are specific claims about specific operations, not a blanket claim that every AI system would use 95% less energy overall. The importance is narrower but still significant: these operations appear in the computational path of AI models, so improving them could affect how efficiently models run.
The researchers tested the approach across language, vision, and reasoning tasks. The tasks included language comprehension, structural reasoning, mathematics, and answering common sense questions. That range suggests the team did not evaluate the algorithm only on a single narrow benchmark type.
- Algorithm: Linear-complexity multiplication, abbreviated as L-Mul.
- Developer: Scientists at BitEnergy AI.
- Reported savings: up to 95% for element-wise floating-point tensor multiplications.
- Dot products: 80% reported energy reduction.
- Test areas: language, vision, and reasoning tasks.
Why Transformer Attention Matters
The researchers say L-Mul can be applied directly to the attention mechanism in transformer models with minimal performance loss. That point is important because attention is a core part of modern language models, including GPT-4o.
Attention mechanisms help transformer models process relationships within data. Because they are central to how these models operate, a change that can be applied there may have a wider impact than a change limited to a less central component.
The source does not claim that L-Mul has already replaced existing systems in production AI models. It says the researchers see a direct path to using the method in attention mechanisms and that the performance loss is minimal. That keeps the claim focused on potential application rather than confirmed deployment.
What BitEnergy AI Wants To Build Next
BitEnergy AI sees broader implications beyond energy efficiency alone. The company believes L-Mul could strengthen academic and economic competitiveness, as well as AI sovereignty. It also sees the method as a way for large organizations to develop custom AI models faster and more cost-effectively.
The next step described by the team is implementation at the hardware level. They also plan to develop programming APIs for high-level model design. That would move L-Mul from an algorithmic idea toward tools and systems that model builders could use more directly.
The stated goal is to train text, symbolic, and multimodal AI models optimized for L-Mul hardware. This matters because software-level efficiency can be limited if the underlying hardware is not designed around the same assumptions. Hardware support could make the algorithm more practical for future AI systems.
The Practical Meaning Of The Claim
The central idea is straightforward: if AI systems can replace expensive mathematical operations with cheaper ones while preserving useful performance, the energy profile of those systems could improve. L-Mul targets that exact tradeoff.
The strongest claims in the source are about operation-level energy savings, not a finished consumer product. The reported figures of up to 95% and 80% are tied to element-wise floating-point tensor multiplications and dot products. The broader promise depends on how well L-Mul can be implemented in hardware and exposed through programming APIs.
Still, the direction is clear. As AI models are used for language comprehension, structural reasoning, mathematics, common sense questions, vision, and multimodal work, energy efficiency becomes part of the core technical challenge. L-Mul offers one route: reduce the cost of repeated computation rather than only trying to make models smaller or systems more powerful.
If BitEnergy AI can carry the method into hardware and model-design tools, L-Mul could become part of how future AI systems are built for lower energy use. For now, the study presents it as a promising algorithmic approach with concrete savings reported for key model operations.