A study published by Nvidia on September 30, 2026 reveals that the average accuracy of AI models drops by 62.8 percent when context length scales from 4,000 to 128,000 tokens. Researchers tested seven open-weight models on simple tasks such as arithmetic, UUID sorting, and table transformation. DeepSeek saw its score drop from 0.909 to 0.554, while Nvidia’s own Nemotron Super fell from 0.711 to 0.138. Missing stable item identifiers caused relative performance drops of up to 64.3 percent. The paper demonstrates that accepting long context and reliably working through it are two different capabilities.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
