Compare AI models

Specifications
Parameters
16B8B/70B
Context window
32k8k
Benchmarks
HumanEval
57.3%81.7%
MATH
27.9%50.4%
MMLU
55.7%82%
GSM8K
72%93%
Overview
CompanyDeepSeekMeta
Release dateMay 16 2024Apr 18 2024
AccessOpen WeightOpen Weight
Model detailsView modelView model

Frequently asked questions

LLaMA 3 (8B/70B) leads DeepSeek V2 Lite on 4 of the 4 benchmarks they both report (HumanEval, MATH, MMLU, GSM8K). LLaMA 3 (8B/70B) shipped 28 days before DeepSeek V2 Lite, so benchmark comparisons should account for the intervening progress.

DeepSeek V2 Lite has 16B parameters, while LLaMA 3 (8B/70B) has 8B/70B. Context windows are 32k (DeepSeek V2 Lite) vs 8k (LLaMA 3 (8B/70B)).

On HumanEval, LLaMA 3 (8B/70B) leads at 81.7% vs DeepSeek V2 Lite at 57.3%. On MATH, LLaMA 3 (8B/70B) leads at 50.4% vs DeepSeek V2 Lite at 27.9%. On MMLU, LLaMA 3 (8B/70B) leads at 82% vs DeepSeek V2 Lite at 55.7%. On GSM8K, LLaMA 3 (8B/70B) leads at 93% vs DeepSeek V2 Lite at 72%.