Two‑Year‑Old Chinese Forecast Gains Global Consensus as Meta, METR and Others Confirm the Same AI Scaling Law

Densing Law of LLMs

Paper "Densing Law of LLMs" (arXiv 2412.04315) was accepted by Nature Machine Intelligence in 2025. The core claim is that model intelligence density grows exponentially, so the number of parameters required for a fixed performance level halves roughly every 3.5 months .

Independent confirmations

Meta scaling ladder : The Muse Spark model released on 2026‑04‑03 achieves the same performance as Llama 4 Maverick (2025) with less than one‑tenth the training compute . When plotted on a common axis the slope matches the density‑law exponential.

METR time‑horizon report : Measuring task duration, the time needed for a model to complete a fixed task doubles every 88.6 days . This time‑scale maps onto the same exponential slope.

Fit quality : Across the original five benchmarks the three datasets produce a fitted line with R²≈0.934 . After filtering with the newly constructed MMLU‑CF dataset the fit improves to R²≈0.953 , indicating an almost perfect exponential relationship.

Implications derived from the law

The law predicts that inference cost should halve every 2.6 months . Real‑world token‑price data from Epoch AI show a 400× drop in a year, with the fastest reductions reaching 900× per year , confirming the prediction.

Because density improves faster than raw parameter scaling, the more sustainable strategy is to increase density through better architectures, higher‑quality data, and smarter training algorithms.

Empirical validation by the MiniCPM series

MiniCPM‑1‑2.4B, released in February 2024, matched or exceeded the performance of Mistral‑7B (September 2023) while using only 35 % of the parameters, a concrete illustration of the halving effect. The MiniCPM family now spans sub‑10 B‑parameter text, multimodal, speech, and full‑modal models and has accumulated over 24 million downloads worldwide.

Key quantitative observations

Parameter‑halving period: 3.5 months.

Training‑compute reduction for equal performance: < 0.1× (Muse Spark vs Llama 4 Maverick).

Task‑duration doubling period: 88.6 days.

Fit R²: 0.934 (original), 0.953 (MMLU‑CF filtered).

Inference‑cost halving period: 2.6 months.

Observed token‑price decline: 400× in one year, up to 900× per year.

References

arXiv: https://arxiv.org/abs/2412.04315

Nature Machine Intelligence: https://www.nature.com/articles/s42256-025-01137-0

Meta blog – Muse Spark: https://ai.meta.com/blog/introducing-muse-spark-msl/

METR – Time Horizon Report: https://metr.org/blog/2026-1-29-time-horizon-1-1/