A researcher has trained a 27-billion-parameter model on a consumer graphics card, reducing the training memory required per parameter from roughly 16 bytes to 1 byte or less. This matters because it could bring large-model training beyond professional data centers and onto personal devices, although it remains well short of low-cost, high-quality training.
What this is
The method targets ternary networks, which represent parameters in one of three states: -1, 0, and 1. This reduces the amount of data that must be stored and updated during training. The traditional AdamW optimizer must maintain multiple pieces of state for every parameter, consuming roughly 16 bytes per parameter; the new method brings that below 1 byte.
In practice, a 27-billion-parameter model can fit on a single 24GB RTX 3090 for training, while a 23-billion-parameter model can be trained on a 16GB graphics card. The author also converted existing Qwen models into a format that can continue training. One 27-billion-parameter version occupies about 8GB and already matches the full-precision version on an existing mathematics test, although it still trails on code-generation tests.
This does not mean “large models can now be trained easily on gaming graphics cards.” A model needs hundreds of billions of training tokens—tokens, or pieces of text—to reach a reasonable level of performance. The current version also exhibits problems such as repetitive answers and forgetting facts. Its code currently permits non-commercial use only, further indicating that the research is still at an early stage.
Industry view
The positive implication is that the infrastructure requirements for local model customization could change. Businesses and researchers have typically needed to secure expensive graphics cards and training frameworks before they can adjust a model’s knowledge or output style. If this method becomes stable, 24GB of memory could become a more widely accessible entry point for experimentation, and models could be deployed more readily on lower-cost devices.
The risks are equally clear. Low-bit training saves memory, but it does not necessarily reduce compute and data costs at the same rate. Consumer graphics cards are slow, and long training runs still incur electricity, cooling, and hardware-wear costs. More importantly, a model’s ability to “remember its training content” is not the same as mastering facts. Current tests focus mainly on mathematics and code, with no large-scale capability evaluation. The tendency to forget facts may make real-world performance worse than the benchmark results suggest.
We believe this work is currently closer to a reproducible experimental path than a mature commercial solution. Whether it evolves from an individual research project into a practical tool depends on cutting training time, gaining more independent validation, and enabling commercial use.
Impact on regular people
For enterprise IT: Small and medium-sized businesses with sensitive-data requirements may gain another way to run local trials. They can inject industry knowledge or test specialized models with less hardware before committing to large-scale training resources. At this stage, however, it is better suited to R&D validation than production-grade model training.
For individual professionals: People with some technical background may use consumer graphics cards to fine-tune models for internal documents, fixed formats, or specialist question-answering. This lowers the practical barrier to data engineering, but training time, data quality, and evaluation expertise remain the primary costs beyond the hardware itself.
For the consumer market: Lighter-weight models may reach more personal computers and edge devices, but training on low-cost graphics cards does not mean consumer AI products will become cheaper overnight. Model stability, factual reliability, and the ability to keep models up to date will affect the user experience more than file size alone.