Atlas

Models

← All models

Llama 3.3 70B Instruct

Meta · 2024-12-06 · 70.6B parameters

Llama 3.3 70B Instruct is Meta's text-only multilingual instruction-tuned model, released on December 6, 2024. It was aligned with supervised fine-tuning and reinforcement learning with human feedback, supports a 128K-token context and eight officially supported languages, and was pretrained on approximately 15 trillion tokens with a December 2023 knowledge cutoff.

Benchmark scores

BenchmarkScore
A-Fantasia69.3
A-Fantasia (Valid Attempts with Retries)69.3
A-Fantasia Backwards Spelling99.0
A-Fantasia Chess34.0
A-Fantasia Chess (Valid Attempts with Retries)34.0
A-Fantasia Cube (Valid Attempts with Retries)75.0
A-Fantasia Cube Rotation75.0
A-Fantasia Spelling (Valid Attempts with Retries)99.0
AA-LCR15.0
AA-LCR v1.115.7
AA-Omniscience Index-54.2
AidanBench776
AIME 2024 (avg@8)24.2
AIME 2024-202516.0
AIME 2025 (Artificial Analysis avg@10)7.7
AIME 2025 (avg@8)7.9
Arabic808
Arena Text — Business, Management & Financial Ops (No Style Control)1243
Arena Text — Chinese (No Style Control)1217
Arena Text — Creative Writing (No Style Control)1250
Arena Text — English (No Style Control)1303
Arena Text — Entertainment, Sports & Media (No Style Control)1241
Arena Text — Exclude Ties (No Style Control)1199
Arena Text — Exclude Ties (Style Controlled)1262
Arena Text — Expert (No Style Control)1225
Arena Text — French (No Style Control)1282
Arena Text — French (Style Controlled)1328
Arena Text — German (No Style Control)1251
Arena Text — Hard Prompts (No Style Control)1257
Arena Text — Hard Prompts English (No Style Control)1277
Loading Atlas data…