Atlas

Models

← All models

Phi-3 Mini 128K Instruct

Microsoft Research · 3.8B parameters

Phi-3 Mini 128K Instruct is the June 2024 updated checkpoint in Microsoft's mutable Phi-3 Mini 128K repository. It is a dense, instruction-tuned 3.8B-class text model trained on a reported 4.9 trillion tokens and post-trained with SFT and DPO; LongRoPE extends its 4,096-position base to a 131,072-token context. The original April 2024 checkpoint remains available at the model card's documented earlier revision.

Benchmark scores

BenchmarkScore
Arena Text — Business, Management & Financial Ops (No Style Control)990
Arena Text — Chinese (No Style Control)1015
Arena Text — Creative Writing (No Style Control)1025
Arena Text — English (No Style Control)1083
Arena Text — Entertainment, Sports & Media (No Style Control)1010
Arena Text — Exclude Ties (No Style Control)845
Arena Text — Exclude Ties (Style Controlled)955
Arena Text — Expert (No Style Control)983
Arena Text — French (No Style Control)1038
Arena Text — French (Style Controlled)1129
Arena Text — German (No Style Control)1006
Arena Text — Hard Prompts (No Style Control)1028
Arena Text — Hard Prompts English (No Style Control)1053
Arena Text — Instruction Following (No Style Control)1023
Arena Text — Japanese (No Style Control)900
Arena Text — Korean (No Style Control)856
Arena Text — Legal & Government (No Style Control)1006
Arena Text — Life, Physical & Social Science (No Style Control)1023
Arena Text — Longer Query (No Style Control)996
Arena Text — Medicine & Healthcare (No Style Control)995
Arena Text — Multi-turn (No Style Control)989
Arena Text — Non-English (No Style Control)1000
Arena Text — Non-English (Style Controlled)1075
Arena Text — Occupational Mathematical (No Style Control)1083
Arena Text — Russian (No Style Control)1003
Arena Text — Software & IT Services (No Style Control)1035
Arena Text — Spanish (No Style Control)1058
Arena Text — Writing, Literature & Language (No Style Control)1040
BBH (Open LLM Leaderboard v2)55.7
BigCodeBench Complete40.6
Loading Atlas data…