Meta

Llama 3.1 70B Instruct

Zero-eval
#1GSM-8K (CoT)
#1MBPP ++ base version
#1MATH (CoT)
+9 more

by Meta

+
+
+
+
About

Llama 3.1 70B was created as a balanced open-source model, designed to provide strong performance with 70 billion parameters while remaining practical for many deployment scenarios. Built to serve as a versatile foundation for fine-tuning and application development, it combines capability with accessibility in the open-source ecosystem.

+
+
+
+
Pricing Range
Input (per 1M)$0.20 -$5.00
Output (per 1M)$0.20 -$10.00
Providers9
+
+
+
+
Timeline
AnnouncedJul 23, 2024
ReleasedJul 23, 2024
+
+
+
+
Specifications
Training Tokens15.0T
+
+
+
+
License & Family
License
Llama 3.1 Community License
Performance Overview
Performance metrics and category breakdown

Overall Performance

18 benchmarks
Average Score
74.7%
Best Score
95.1%
High Performers (80%+)
10

Performance Metrics

Max Context Window
256.0K
Avg Throughput
213.4 tok/s
Avg Latency
0ms
+
+
+
+
All Benchmark Results for Llama 3.1 70B Instruct
Complete list of benchmark scores with detailed information
GSM-8K (CoT)
text
0.95
95.1%
Self-reported
ARC-C
text
0.95
94.8%
Self-reported
API-Bank
text
0.90
90.0%
Self-reported
IFEval
text
0.88
87.5%
Self-reported
Multilingual MGSM (CoT)
text
0.87
86.9%
Self-reported
MBPP ++ base version
text
0.86
86.0%
Self-reported
MMLU (CoT)
text
0.86
86.0%
Self-reported
BFCL
text
0.85
84.8%
Self-reported
MMLU
text
0.84
83.6%
Self-reported
HumanEval
text
0.81
80.5%
Self-reported
Showing 1 to 10 of 18 benchmarks
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+