How quickly do open models catch up to frontier AI?

Epoch AI researchers Venkat Somala and Luke Emberson report that "anyone can locally run models matching the absolute frontier of LLM performance from just 6 to 12 months ago" using a single top-of-the-line consumer gaming GPU such as Nvidia's RTX 5090, which costs "under $2500." The finding sharpens Epoch AI's earlier, broader estimate of "a 5 to 22 month gap for open-weight models of any size" into a narrower range specific to models sized to fit entirely on one consumer card. The comparison runs leading open models constrained to roughly 28B parameters (the RTX 4090 era) or 40B parameters (the RTX 5090 era) against the closed-source frontier across four measures: GPQA-Diamond, MMLU-Pro, the Artificial Analysis Intelligence Index, and LM Arena Elo scores.

Three factors drive the narrowness of the lag, per the excerpt: "a comparable rate of scaling among open-weight models to the closed-source frontier," "the success of techniques like model distillation," and "continual progress in GPUs enabling larger models to be run at home." The hardware ceiling itself is quantified rather than assumed: at 4-bit quantization (about 0.54 bytes per parameter) plus an 8k-context KV cache (about 6.6 GiB) and roughly 2 GiB of runtime overhead, Epoch AI puts the RTX 4090's theoretical ceiling at "~27B parameters" and the RTX 5090's at "~40B parameters."

This sits alongside Can small language models handle most agent tasks? and What actually limits language models on mobile phones?: those notes argue hardware and task structure push deployment toward smaller models regardless of quality preference, and this finding adds that the smaller models forced onto consumer hardware are not far behind frontier capability — the lag is measured in months, not generations. It also connects to Can frontier exams really measure cutting-edge AI capability?: Epoch AI picks GPQA-Diamond and MMLU-Pro partly because they "have remained unsaturated over a reasonably long period," the same saturation concern that motivates harder exams elsewhere.

The excerpt itself qualifies its own measurement: "small open models are more likely to be optimized for specific benchmarks, so the real-world lag may be somewhat longer" than the benchmark-measured 6 to 12 months. It does not establish whether this lag will keep narrowing, hold steady, or widen as frontier models grow larger and harder to distill, and it says nothing about inference speed, context length, or agentic reliability at consumer-GPU scale — only benchmark accuracy and preference scores. If the pattern holds, it implies that access to near-frontier capability is becoming less bottlenecked by cost or hardware, independent of whatever access restrictions model providers impose by other means.