The Memo - Special edition - Meta AI Llama 4 Behemoth 2T on ~30T tokens - Apr/2025

Llama 4 Behemoth outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro.

Apr 05, 2025

To:      US Govt, major govts, Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, Tesla, Citi, Tencent, IBM, & 10,000+ more recipients…
From:    Dr Alan D. Thompson <LifeArchitect.ai>
Sent:    5/Apr/2025
Subject: The Memo - AI that matters, as it happens, in plain English
AGI:     92% ➜ 94%
ASI:     0/50 (no expected movement until post-AGI)

Meta announces Llama 4 Behemoth 2T on ~30T tokens

Once again, we have this out to The Memo readers within just a few hours (actually minutes) of model release.

Wow! Meta AI has trained a 2 trillion parameter MoE ‘teacher model, Llama 4 Behemoth, that outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks such as MATH-500 and GPQA Diamond.‘

In plain English: Meta AI’s Llama models are famously open source (available for free), and have been downloaded hundreds of millions of times. The latest version, Llama 4, comes about a year after Llama 3. It now boasts 10 million tokens of context (roughly equal to the number of words in and out), allowing you to give it complete access to a personal library during conversation, or to have it generate 10 million words while keeping all of that info in its short-term memory. It is multimodal, meaning that it was trained on ‘unlabeled text, image, and video data,’ and can respond with any of these. The 2T model, Behemoth, would be the largest publicly known model in the world.