What is Llama 4 Behemoth?
Llama 4 Behemoth is Meta's most powerful AI model — a massive 2 trillion parameter system with 288 billion active parameters across 16 experts. Announced alongside Scout and Maverick in April 2025, Behemoth has not yet been publicly released as open weights. It currently serves as a "teacher model" used to improve Scout and Maverick through codistillation — Behemoth's knowledge is distilled into the smaller models to boost their performance without adding parameters.
Behemoth's Known Performance
Based on Meta's benchmarks at announcement, Behemoth outperformed GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks including MATH-500 and GPQA Diamond. It fell behind OpenAI's o1, DeepSeek-R1, and Gemini 2.5 Pro on the hardest reasoning tasks, suggesting it's a very capable general model rather than a specialised reasoning model like o1.
Behemoth was trained using 32,000 GPUs in FP8 precision, achieving 390 TFLOPs per GPU — a landmark for training efficiency at this scale. Meta used asynchronous reinforcement learning, curriculum sampling based on prompt difficulty, and a new distillation loss function to train it effectively.
When Will Behemoth Be Released?
As of May 2026, Meta has not announced a specific release date for Behemoth's public weights. Meta indicated it is "still training" at the April 2025 announcement, and subsequent updates have not confirmed a release timeline. The model may eventually be released as open weights, made available via API only, or used exclusively as an internal teacher model. Meta has signalled it intends to release Behemoth, but the timeline remains unclear.
Why Does Behemoth Matter?
Even without public release, Behemoth's existence raises the ceiling for open-source AI. Every improvement Meta makes to Behemoth can be distilled into future Scout and Maverick updates, meaning Llama 4's descendants will benefit from Behemoth's capabilities without requiring users to run a 2 trillion parameter model. It demonstrates that open-source labs can train models at the same scale as closed-source leaders.
Frequently Asked Questions
Can I use Llama 4 Behemoth today?
No. Behemoth is not publicly available as of May 2026. Only Scout and Maverick are released. Meta AI's consumer assistant across WhatsApp, Facebook, and Instagram uses Llama 4 models, possibly including Behemoth internally, but developer access is not available.
How big is Llama 4 Behemoth?
Behemoth has approximately 2 trillion total parameters with 288 billion active per inference pass, organised through 16 experts in a MoE architecture. It requires multiple nodes of high-end GPUs to run inference.