Model
Model explorer

Llama 3.1 405B

OPEN
Meta · Llama 3.1 family · released Jul 23, 2024

First openly available model rivaling top closed frontier models on knowledge, math, and tool use; 128K context.

ReasoningCodingVisionFunction callingTool useAgentic
1237.4
Elo · rank #173
Parameters
405B
Active params
405B (dense)
Context
128K tokens
Architecture
Dense decoder-only Transformer (GQA)
License
Llama 3.1 Community License Agreement
Languages
8+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-ChallengeReasoning96.9%#1
best: this model · 96.9%
Arena EloHuman preference1229#38
best: Claude Fable 5 · 1505
BFCLAgents88.5%#5
best: SmolLM3 3B (No Thinking) · 92.3%
GPQA DiamondReasoning50.7%#187
best: GPT-5.6 · 94.6%
GSM8KMath96.8%#1
best: this model · 96.8%
HumanEval+Coding82.3%#13
best: Mistral Small 3.2 (24B) · 92.9%
HumanEvalCoding89.0%#25
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning88.6%#33
best: Gemma 4 26B A4B · 98.5%
LogicKorHuman preference8.79#8
best: GPT-4o · 9.33
MATH-500Math73.8%#82
best: GPT-5 · 99.4%
MBPP+Coding88.6%#1
best: this model · 88.6%
MBPPCoding78.8%#17
best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%
MGSMMath91.6%#6
best: OpenAI o4-mini · 93.7%
MMLU-ProKnowledge73.3%#83
best: Claude Fable 5 · 91.5%
MMLUKnowledge87.3%#35
best: OpenAI o3 · 92.9%
Run it locally
VRAM @ Q4
230 GB
VRAM @ FP16
810 GB
Fits on (Q4)
M3 Ultra 512GB
Far exceeds a single RTX 4090's 24GB VRAM even at Q4 (~230GB); needs multi-GPU or heavy CPU/disk offload, so no meaningful single-4090 tok/s figure is reported
Quantizations
GGUF Q4 · AWQ · GPTQ · MLX
Llama 3.1 family
Elo progression across releases
Llama 3.1 405BJul 20241237.4Llama 3.1 70BJul 2024980.8Llama 3.1 8BJul 2024553.8
Resources
HFmeta-llama/Meta-Llama-3.1-405B-InstructGHmeta-llama/llama-modelsDOCSllama.com/docs/model-cards-and-prompt-formats/llama3_1APIAvailable via Meta API
API price weights · each benchmark row carries its own source badge (see methodology)