Microsoft says it completed training of a 2.8-trillion-parameter AI model, touting efficiency gains

Microsoft has completed pre-training for a new large AI model with a reported total size of 2.8 trillion parameters, according to a market report citing company disclosures. Parameters are the learned numerical weights inside a model, and the count is often used as a rough proxy for model capacity and potential performance, though it does not fully determine real-world usefulness.

In the report, Microsoft claims the model delivered stronger results than every open-weight model it tested. Open-weight models are AI systems whose model weights are publicly released, enabling other teams to run, inspect, or fine-tune them without needing the original provider’s closed infrastructure.

Microsoft also told the publication that the new system required up to 12 times less computing power than those benchmarks, framing the improvement as an efficiency gain rather than only a raw performance upgrade. The report describes the difference as being tied to how much compute is needed to reach comparable or better results.

The disclosure is part of the broader push by major cloud and software providers to expand their AI capabilities for enterprise and developer customers. Microsoft’s AI work typically ties into its Azure cloud platform, where model training, tuning, and deployment are bundled with managed services for customers building applications that use generative AI.

For the market, claims of both higher accuracy relative to open-weight rivals and lower compute requirements can be significant because AI costs are sensitive to training and iteration. Compute refers to the processing time and specialized hardware cycles needed to train or refine models. If Microsoft’s efficiency claims hold up in independently verifiable tests, they could translate into lower operating costs and more scalable experimentation for customers.

Still, the information in the market report leaves several items unclear. It does not specify the exact model name, the training timeline, or the detailed evaluation methodology used to compare against open-weight models. It also does not provide the specific tasks, datasets, or performance metrics used to support the statement that the new model outperformed every open-weight model it tested.

Microsoft did not appear to disclose in the report how the “up to 12 times less computing power” figure was calculated, such as whether it refers to total training compute, compute to reach a particular performance threshold, or compute relative to a particular set of baselines. Without those details, the magnitude of the efficiency gain and its reproducibility may be difficult for outside researchers to validate.

What to watch next is whether Microsoft publishes more technical documentation, evaluation results, or model-release plans that would allow developers to compare performance and cost characteristics more directly. For investors and customers monitoring the AI arms race, the key question is not only how large the model is, but how consistently its reported efficiency and benchmark advantages show up across standard tests and real deployment scenarios.