Atlas

Models

← All models

Llama 3.1 70B Instruct

Meta · 2024-07-23 · 70.6B parameters

Llama 3.1 70B Instruct is Meta's instruction-tuned, text-only 70B-class model from the July 2024 Llama 3.1 release, with multilingual dialogue, tool use, and a 131,072-token context window. Meta reports roughly 15 trillion pretraining tokens and a December 2023 data cutoff; exact cumulative exposure including post-training is not disclosed. The open weights are released under the Llama 3.1 Community License.

Benchmark scores

BenchmarkScore
AA-LCR6.3
AA-Omniscience Index-40.9
Agentic Tool Use (Chat)33.5
Agentic Tool Use (Enterprise)37.2
AidanBench578
AIME 20254.0
Arena Text — Business, Management & Financial Ops (No Style Control)1235
Arena Text — Chinese (No Style Control)1214
Arena Text — Creative Writing (No Style Control)1232
Arena Text — English (No Style Control)1294
Arena Text — Entertainment, Sports & Media (No Style Control)1230
Arena Text — Exclude Ties (No Style Control)1177
Arena Text — Exclude Ties (Style Controlled)1222
Arena Text — Expert (No Style Control)1208
Arena Text — French (No Style Control)1260
Arena Text — French (Style Controlled)1300
Arena Text — German (No Style Control)1222
Arena Text — Hard Prompts (No Style Control)1241
Arena Text — Hard Prompts English (No Style Control)1263
Arena Text — Instruction Following (No Style Control)1232
Arena Text — Japanese (No Style Control)1133
Arena Text — Korean (No Style Control)1141
Arena Text — Legal & Government (No Style Control)1283
Arena Text — Life, Physical & Social Science (No Style Control)1276
Arena Text — Longer Query (No Style Control)1241
Arena Text — Medicine & Healthcare (No Style Control)1256
Arena Text — Multi-turn (No Style Control)1256
Arena Text — Non-English (No Style Control)1219
Arena Text — Non-English (Style Controlled)1257
Arena Text — Occupational Mathematical (No Style Control)1254
Loading Atlas data…