Consumer Gpu Model Gap

Epoch AI asked how long it takes for frontier-level benchmark scores to become reachable on hardware an individual can buy. Its August 2025 data insight finds that a single top gaming GPU, such as NVIDIA's RTX 5090 at under $2,500, can run open models that match the best scores of 6 to 12 months earlier. That fits Epoch's earlier estimate that open-weight models of any size trail the frontier by 5 to 22 months.

The definition of runnable at home does most of the work. Epoch included only models whose full weights fit in the card's memory when compressed to 4 bits per weight, plus room for an 8,000-token context and runtime overhead. That caps models at about 28 billion parameters for the RTX 4090 era and 40 billion for the RTX 5090 era. The size of the gap then depends on the yardstick: about 6.3 months on the Artificial Analysis Intelligence Index, 7.3 on MMLU-Pro, 7.4 on GPQA-Diamond and 12.4 on LM Arena.

Epoch flags several reasons the practical gap may be wider. Small open models are more likely to be optimized for specific benchmarks, so real-world lag may be somewhat longer. Requiring the whole model to fit in memory biases the sample against mixture-of-experts models, which activate only part of their parameters at a time. The 8,000-token assumption also likely understates the memory needed for long outputs, especially from reasoning models, while offloading or heavier compression could run larger models at some cost to speed or quality.