Llama 4 Behemoth

// Overview

Llama 4 Behemoth is the largest model in Meta's Llama 4 "herd," unveiled alongside Llama 4 Scout and Maverick on April 5, 2025. Meta describes it as a natively multimodal mixture-of-experts (MoE) model with 288 billion active parameters drawn from a total parameter count approaching 2 trillion, organized across 16 experts. It accepts text and image input and was positioned as one of the most capable base models in the world for non-reasoning tasks.

Behemoth's primary published role is as a teacher model. Meta codistilled the smaller Llama 4 Maverick from Behemoth using a novel distillation loss that dynamically weights soft and hard targets during training, which Meta credits for substantial quality gains in the shipped models. On STEM benchmarks, Meta reported that Behemoth outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several measures including MATH-500 and GPQA Diamond.

Crucially, Llama 4 Behemoth was never publicly released. At the April 2025 launch Meta stated it was "not yet releasing Llama 4 Behemoth as it is still training." Through 2025 and into 2026 its launch was repeatedly delayed amid reports that internal teams were unsure the gains justified shipping a 2-trillion-parameter model, leaving it effectively shelved — never formally cancelled, but with no public weights, API, or pricing.

// Benchmarks

1. MATH-50095%
2. MMLU Pro82.2%
3. GPQA Diamond73.7%
4. MMMU76.1%
5. Multilingual MMLU85.8%
6. LiveCodeBench49.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• State-of-the-art reported results for a non-reasoning model on math, multilinguality, and image benchmarks at the April 2025 announcement
• Outperformed GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM benchmarks per Meta
• Native multimodality (text + image) via an early-fusion MoE backbone
• Highly effective as a teacher model: codistilled Maverick for substantial downstream quality gains

// Best for

• Knowledge distillation — serving as a teacher to train smaller, deployable Llama 4 models (Scout, Maverick)
• Internal research benchmark for frontier-scale non-reasoning multimodal performance
• Reference point for evaluating large-scale MoE design and codistillation approaches

// Llama 4 — every version

The full lineage of the Llama 4 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.