Atlas

Models

← All models

Phi-4

Microsoft Research · 2024-12-12 · 14.7B parameters

Phi-4 is Microsoft's 14B-parameter small language model (SLM) released in December 2024, emphasizing high-quality synthetic training data to achieve strong performance in complex reasoning, mathematics, and coding. It was trained using multi-agent, self-revision synthetic data pipelines and fine-tuned with SFT and DPO alignment. Despite its compact size, Phi-4 is competitive with significantly larger models on reasoning-heavy tasks.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-56.7
AIME 202518.0
Arena Text — Business, Management & Financial Ops (No Style Control)1199
Arena Text — Chinese (No Style Control)1211
Arena Text — Creative Writing (No Style Control)1182
Arena Text — English (No Style Control)1231
Arena Text — Entertainment, Sports & Media (No Style Control)1163
Arena Text — Exclude Ties (No Style Control)1101
Arena Text — Exclude Ties (Style Controlled)1157
Arena Text — Expert (No Style Control)1202
Arena Text — French (No Style Control)1223
Arena Text — French (Style Controlled)1265
Arena Text — German (No Style Control)1223
Arena Text — Hard Prompts (No Style Control)1220
Arena Text — Hard Prompts English (No Style Control)1231
Arena Text — Instruction Following (No Style Control)1202
Arena Text — Japanese (No Style Control)1159
Arena Text — Korean (No Style Control)1151
Arena Text — Legal & Government (No Style Control)1241
Arena Text — Life, Physical & Social Science (No Style Control)1207
Arena Text — Longer Query (No Style Control)1217
Arena Text — Medicine & Healthcare (No Style Control)1201
Arena Text — Multi-turn (No Style Control)1206
Arena Text — Non-English (No Style Control)1197
Arena Text — Non-English (Style Controlled)1233
Arena Text — Occupational Mathematical (No Style Control)1254
Arena Text — Russian (No Style Control)1208
Arena Text — Software & IT Services (No Style Control)1229
Arena Text — Spanish (No Style Control)1231
Loading Atlas data…