Model

Llama 3 70B

OPEN

Meta · Llama 3 family · released Apr 18, 2024

Llama 3 flagship; Instruct checkpoint scored 82.0% MMLU and 81.7% HumanEval at launch, competitive with Gemini 1.5 Pro and beating Claude 3 Sonnet on several benchmarks. Officially English-only at this release (multilingual support came with Llama 3.1).

ReasoningCodingVisionFunction callingTool useAgentic

886.4

Elo · rank #239

Parameters

70B

Active params

70B (dense)

Context

8K tokens

Architecture

Dense transformer decoder with grouped-query attention (GQA), 128K-token vocabulary

License

Llama 3 Community License Agreement

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning93.0%#17

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference41.1%#46

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

Arena EloHuman preference1206#40

best: Claude Fable 5 · 1505

BIG-Bench HardReasoning81.3%#31

best: ERNIE 4.5 300B-A47B · 94.3%

DROPReasoning79.7#27

best: Hunyuan-T1 · 93.1

GPQA DiamondReasoning39.5%#223

best: GPT-5.6 · 94.6%

GSM8KMath93.0%#22

best: Llama 3.1 405B · 96.8%

HumanEvalCoding81.7%#51

best: Claude Opus 4.5 · 99.4%

MATH-500Math50.4%#131

best: GPT-5 · 99.4%

MMLUKnowledge79.5%#95

best: OpenAI o3 · 92.9%

TriviaQAKnowledge89.7%#2

best: Sarvam-1 (2B) · 90.6%

WinoGrandeReasoning83.1%#18

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

40 GB

VRAM @ FP16

140 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

~18 tok/s on RTX 4090 (Q4, llama.cpp)

Quantizations

GGUF Q4 · AWQ · GPTQ · MLX

Llama 3 family

Elo progression across releases

Llama 3 70BApr 2024886.4 Llama 3 8BApr 2024486.3

Resources

HFmeta-llama/Meta-Llama-3-70B-Instruct GHmeta-llama/llama3 DOCSllama.meta.com APIAvailable via Meta API

API price weights · each benchmark row carries its own source badge (see methodology)