Skip to main content
llm.info

Llama 3.3 70B

Open Weights

Meta

Meta's efficiency-optimized model delivering Llama 3.1 405B-level quality at fraction of the cost. Released December 2024 with breakthrough cost-performance ratio - generates responses nearly 5x more cost-efficiently than 405B while maintaining similar output quality. Features 128K context window and achieves 86% on MMLU Chat, matching 3.1 70B. Excels at math (77% MATH, up from 67.8%), multilingual reasoning (91.1% MGSM vs 86.9%), and instruction following (92.1% IFEval). Pretrained on 15T tokens using 39.3M H100 GPU hours. Runs on accessible hardware (2-4 A100s).

Strengths

  • Matches Llama 3.1 405B quality at nearly 5x better cost-efficiency
  • Dramatic math improvement - 77% MATH score (up from 67.8% in 3.1 70B)
  • Outstanding multilingual reasoning - 91.1% MGSM (up from 86.9%)
  • Excellent instruction following - 92.1% IFEval score
  • 128K context window handles long documents
  • More accessible hardware requirements than 405B (2-4 A100s vs 8+)
  • Permissive license enables commercial use and fine-tuning

Caveats

  • No vision or audio capabilities (text-only)
  • December 2023 training cutoff older than some 2024-2025 releases
  • 4K max output tokens lower than newer models
  • Still requires significant resources for self-hosting (140GB+ VRAM)
  • API pricing varies across providers ($0.40-$0.88 per 1M)
Vendor
Meta
Context Window
128,000 tokens
Max Output
4,096 tokens
License
Llama 3.3 Community License
Release Date
2024-12-06
Training Cutoff
2023-12-01
Modalities
text
BenchmarkScore
MMLU Chat86.0
MATH77.0
MGSM91.1
IFEval92.1

Capabilities

Vision
Audio
Video
Tool Use

Pricing

Open weights - free to download. API providers typically $0.40-$0.88 per 1M input tokens

Reviews

Comments