Atlas

Models

← All models

Llama 4 Maverick Instruct FP8

Meta · 2025-04-05 · 401.6B parameters

Llama 4 Maverick Instruct FP8 is Meta's separately released FP8-quantized checkpoint of the post-trained, natively multimodal Llama 4 Maverick Instruct model. It preserves the 128-expert architecture, 1,048,576-token context window, and text-and-image inputs of the BF16 checkpoint; the released safetensors expose 401,649,841,664 tensor elements across BF16 and FP8 weights.

Benchmark scores

BenchmarkScore
ALE-Bench173
ARC-AGI-14.4
ARC-AGI-20.0
BFCL v4 Overall Accuracy37.3
Capability131.7
ForecastBench Baseline Dataset Brier Index57.2
ForecastBench Baseline Market Brier Index57.0
ForecastBench Baseline Overall Brier Index57.1
ForecastBench Preliminary Dataset Brier Index57.1
ForecastBench Tournament Dataset Brier Index57.2
ForecastBench Tournament Market Brier Index68.6
ForecastBench Tournament Overall Brier Index62.5
GPQA Diamond67.0
MASK49.7
MATH Level 573.0
OTIS Mock AIME 2024-202520.6
Loading Atlas data…