Compare AI models

Specifications
Parameters
8B/70B/405B16B
Context window
128k8k
API pricing
Cheapest input
$0.02DeepInfra—
Cheapest output
$0.04DeepInfra—
$6.00DeepInfra · fp8—
Benchmarks
HumanEval
89%48.1%
MATH
73.8%45.3%
GPQA Diamond
49%35.2%
MMLU-Pro
73.3%42.4%
MMLU
87.3%70%
GSM8K
96.8%77.4%
Overview
CompanyMetaMoonshot AI
Release dateJul 23 2024Feb 22 2025
AccessOpen WeightOpen Weight
Model detailsView modelView model

Frequently asked questions

LLaMA 3.1 leads Moonlight on 6 of the 6 benchmarks they both report. LLaMA 3.1 shipped 214 days before Moonlight, so benchmark comparisons should account for the intervening progress.

LLaMA 3.1 has 8B/70B/405B parameters, while Moonlight has 16B. Context windows are 128k (LLaMA 3.1) vs 8k (Moonlight).

On HumanEval, LLaMA 3.1 leads at 89% vs Moonlight at 48.1%. On MATH, LLaMA 3.1 leads at 73.8% vs Moonlight at 45.3%. On GPQA Diamond, LLaMA 3.1 leads at 49% vs Moonlight at 35.2%. On MMLU-Pro, LLaMA 3.1 leads at 73.3% vs Moonlight at 42.4%. On MMLU, LLaMA 3.1 leads at 87.3% vs Moonlight at 70%. On GSM8K, LLaMA 3.1 leads at 96.8% vs Moonlight at 77.4%.