Note: Overall leaderboard rankings may not reflect true model quality — individual benchmarks give a clearer picture. ARC-Challenge MMLU GPQA GSM8K Artificial Analysis Intelligence Index v4.1
← Back to leaderboard

Microsoft: phi-4-mini-instruct

17 benchmarks
CompanyMicrosoft
Country of originUSA
WeightsOpen weights
Parameter count4.0B
Architecturephi3
VariantReasoning
Tensor typebf16
Release year2025
OriginInstruction tuning

Leaderboard position

#814
Overall rank (of 1121 models)
434
Elo rating
17
Benchmarks
Top 72.6%
Percentile
#11
In Microsoft (of 19 models)
Above: Apple: diffucoder-7b-instruct #813 Below: Moonshot AI: Moonlight Instruct #815

Benchmark highlights

Best benchmark
ARC-Easy 77.5439
Worst benchmark
WikiText-2 (-ppl) -10.8412

Nearest competitor from another company

The model from another company with the most similar benchmark scores.

appeared as nearest in 2 of 17 benchmarks

Technology Innovation Institute: falcon-h1-1.5b-instruct
2/17 11.8%
Qwen: qwen3.5-9b-base
2/17 11.8%
Model Votes % of benchmarks
Technology Innovation Institute: falcon-h1-1.5b-instruct 2 11.8%
Qwen: qwen3.5-9b-base 2 11.8%
Meta Llama: llama-2-70b-chat 1 5.9%
Meta Llama: Llama 3.2 3B Instruct 1 5.9%
Google: gemma-3-12b-pt 1 5.9%
DeepSeek: deepseek-r1-distill-qwen-1.5b 1 5.9%
Qwen: Qwen3.5 0.8B 1 5.9%
Meta Llama: Llama 3.3 70B Instruct 1 5.9%
Apple: openelm-1_1b-instruct 1 5.9%
Qwen: Qwen3.5 9B 1 5.9%

Benchmarks

ARC-Easy 77.5439 WinoGrande 71.9021 HellaSwag 71.5 MATH-500 69.6 ARC-Challenge 54.1806 CowQA-RU-Rapid 50.93 MMLU-Pro 46.5 MMLU 39.0181 GPQA Diamond 33.1 TruthfulQA 32.0685 IFBench 21.1 Artificial Analysis Intelligence Index v4.0 (Maximum Reasoning) 8.4 AIME 2025 6.7 Humanity's Last Exam (no tools) 4.2 Artificial Analysis Intelligence Index v4.1 (Maximum Reasoning) 3 -(CoderPPL) -2.4022 WikiText-2 (-ppl) -10.8412

Predicted scores

Estimated from 25 similar models and the global Elo rating. These are projections, not measured results.

Benchmark Predicted Confidence Based on
IFEval ~76.05 67% Google: Gemma 3 12B Instruct, Google: Gemma 3 27B Instruct, Meta Llama: Llama 3.2 3B Instruct
BBH ~50.21 67% Meta Llama: llama-2-7b-chat, Meta Llama: llama-2-13b-chat, DeepSeek: deepseek-r1-distill-qwen-1.5b
GSM8K ~65.1 66% IBM Granite: granite-4.1-3b, Qwen: qwen2.5-coder-7b-instruct, Hugging Face Smol Models Research: smollm3-3b-base
HumanEval ~60.62 65% DeepSeek: deepseek-r1-distill-qwen-1.5b, Qwen: qwen-1.8b, OpenAI: GPT-3.5 Turbo
MBPP ~55.98 65% DeepSeek: deepseek-r1-distill-qwen-1.5b, Qwen: qwen-1.8b, Qwen: qwen3-0.6b-base
CowQA ~51.68 65% Meta Llama: Llama 3.2 3B Instruct, IBM Granite: Granite 4.1 8B, MWS AI: cotype-nano
GPQA ~33.44 65% Meta Llama: Llama 3.2 3B Instruct, Qwen: qwen3-0.6b-base, LiquidAI: lfm-2-8b-a1b
SuperGPQA ~29.69 64% Qwen: qwen2.5-coder-7b-instruct, Qwen: qwen3-0.6b-base, Qwen: qwen2.5-0.5b
HumanEval+ ~67.79 62% Qwen: qwen2.5-coder-7b-instruct, DeepSeek: deepseek-r1-distill-qwen-1.5b, Technology Innovation Institute: falcon-h1-1.5b-deep-base
MBPP+ ~62.54 62% Qwen: qwen2.5-coder-7b-instruct, DeepSeek: deepseek-r1-distill-qwen-1.5b, Technology Innovation Institute: falcon-h1-1.5b-deep-base
MATH ~40.35 62% Meta Llama: llama-2-7b-chat, Meta Llama: llama-2-13b-chat, Qwen: qwen-1.8b
MMLU-Redux ~62.81 62% DeepSeek: deepseek-r1-distill-qwen-1.5b, Qwen: qwen3-0.6b-base, Qwen: qwen3-1.7b
PIQA ~78.69 61% Qwen: qwen-1.8b, Google: gemma-3-4b-pt, Apple: openelm-3b-instruct
Text Arena (Snapshot from 20 june 2026) ~1266.72 61% Google: Gemma 3 12B Instruct, Meta Llama: llama-2-13b-chat, OpenAI: GPT-3.5 Turbo
SciCode ~19.49 60% Google: Gemini 1.0 Pro, Cohere Labs: Tiny Aya Global, Meta Llama: Llama 3.3 70B Instruct
Artificial Analysis Intelligence Index v5.0.1 ~12 60% Qwen: qwen3.8-27b
LiveCodeBench v6 ~42.85 60% DeepSeek: deepseek-r1-distill-qwen-1.5b, Qwen: qwen3-1.7b, Hugging Face Smol Models Research: smollm3-3b
Artificial Analysis Intelligence Index v4.3.2 ~7.36 60% Qwen: Qwen3.5 2B, LiquidAI: LFM2.5 1.2B Instruct, Qwen: Qwen3.5 9B
SWE-bench Verified ~25.32 59% Qwen: qwen2.5-coder-7b-instruct, Qwen: Qwen3 8B, Qwen: Qwen3 4B
AA-LCR ~33.89 59% Cohere Labs: Tiny Aya Global, Meta Llama: Llama 3.3 70B Instruct, LiquidAI: lfm-3-1.7b
LLMArena (llmarena.ru) (Snapshot from 19 june 2026) ~955.64 59% T-Tech: t-lite-it-1.0, MWS AI: cotype-nano, Meta Llama: Llama 3.3 70B Instruct
SWE-bench Pro ~7.72 58% Google: Gemma 3 27B Instruct, Meta Llama: Llama 4 Maverick Instruct, Meta Llama: Llama 3.1 405B Instruct
Artificial Analysis Agentic Index (Maximum Reasoning) ~6.55 58% Google: Gemma 4 E2B Instruct, Qwen: Qwen3 0.6B, Meta Llama: Llama 3.3 70B Instruct
Humanity's Last Exam (with tools) ~11.9 57% OpenAI: gpt-oss-20b, LiquidAI: lfm-3-1.7b, Google: diffusiongemma-26b-a4b-it
MMMU-Pro ~48.1 57% Google: Gemma 4 E4B Instruct, Google: diffusiongemma-26b-a4b-it, Google: Gemma 4 12B Instruct