Atlas

Models

← All models

Llama 3.2 3B Instruct

Meta · 2024-09-25 · 3.2B parameters

Llama 3.2 3B Instruct is Meta's 3-billion-parameter instruction-tuned language model from the Llama 3.2 family, optimized for conversational and agentic use cases on edge and mobile devices. It is a dense decoder-only Transformer with a 128K context window, post-trained through several rounds of supervised fine-tuning, rejection sampling, and direct preference optimization. Designed for efficient on-device deployment, it supports tasks such as summarization, question answering, and tool use at low memory and compute cost.

Benchmark scores

BenchmarkScore
AA-LCR2.0
AidanBench147
AIME 20253.3
Arena Text — Business, Management & Financial Ops (No Style Control)1073
Arena Text — Chinese (No Style Control)1016
Arena Text — Creative Writing (No Style Control)1095
Arena Text — English (No Style Control)1180
Arena Text — Entertainment, Sports & Media (No Style Control)1084
Arena Text — Exclude Ties (No Style Control)934
Arena Text — Exclude Ties (Style Controlled)1009
Arena Text — Expert (No Style Control)1089
Arena Text — German (No Style Control)1056
Arena Text — Hard Prompts (No Style Control)1095
Arena Text — Hard Prompts English (No Style Control)1143
Arena Text — Instruction Following (No Style Control)1089
Arena Text — Legal & Government (No Style Control)1115
Arena Text — Life, Physical & Social Science (No Style Control)1106
Arena Text — Longer Query (No Style Control)1101
Arena Text — Medicine & Healthcare (No Style Control)1103
Arena Text — Multi-turn (No Style Control)1105
Arena Text — Non-English (No Style Control)1019
Arena Text — Non-English (Style Controlled)1088
Arena Text — Occupational Mathematical (No Style Control)1129
Arena Text — Russian (No Style Control)948
Arena Text — Software & IT Services (No Style Control)1104
Arena Text — Writing, Literature & Language (No Style Control)1096
Artificial Analysis Intelligence Index4.2
BALROG10.1
BBH (Open LLM Leaderboard v2)46.1
BFCL v4 Overall Accuracy21.9
Loading Atlas data…