Model

Phi-3.5-MoE (16x3.8B, 6.6B active)

OPEN

Microsoft · Phi-3.5 family · released Aug 20, 2024

First Phi mixture-of-experts (16 experts, 6.6B active); rivals Gemini-1.5-Flash and beats Llama-3.1 8B, MIT-licensed.

ReasoningCodingVisionFunction callingTool useAgentic

791.5

Elo · rank #254

Parameters

41.9B

Active params

6.6B (MoE)

Context

128K tokens

Architecture

Mixture-of-Experts decoder-only Transformer (16 experts, 2 active per token)

License

MIT

Languages

22+

API price (in/out)

$0.13 / $0.52

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning91.0%#23

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference37.9%#47

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

BIG-Bench HardReasoning79.1%#35

best: ERNIE 4.5 300B-A47B · 94.3%

GPQA DiamondReasoning36.8%#234

best: GPT-5.6 · 94.6%

GSM8KMath88.7%#48

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning83.8%#55

best: Claude 3 Opus · 95.4%

HumanEvalCoding70.7%#84

best: Claude Opus 4.5 · 99.4%

MATH-500Math59.5%#113

best: GPT-5 · 99.4%

MBPPCoding80.8%#12

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MGSMMath58.7%#41

best: OpenAI o4-mini · 93.7%

MMLU-ProKnowledge54.3%#125

best: Claude Fable 5 · 91.5%

MMLUKnowledge78.9%#100

best: OpenAI o3 · 92.9%

MMMLUKnowledge69.9%#41

best: Gemini 3.1 Pro · 92.6%

OpenBookQAReasoning89.6%#2

best: Claude 1 · 90.8%

PIQAReasoning88.6%#3

best: GPT-4o mini · 93.1%

Social IQaReasoning78.0%#5

best: Apple DCLM-Baseline 7B · 82.9%

TruthfulQAKnowledge77.5%#1

best: this model · 77.5%

WinoGrandeReasoning81.3%#30

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

27 GB

VRAM @ FP16

85 GB

Fits on (Q4)

RTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF Q4

Phi-3.5 family

Elo progression across releases

Phi-3.5-mini (3.8B)Aug 2024477.1 Phi-3.5-MoE (16x3.8B, 6.6B active)Aug 2024791.5 Phi-3.5-vision (4.2B)Aug 2024532.4

Resources

HFmicrosoft/Phi-3.5-MoE-instruct GHmicrosoft/PhiCookBook DOCStechcommunity.microsoft.com/blog/azure-ai-foundry-blog/discover-the-new-multi-lingual-high-quality-phi-3-5-slms/4225280 APIAvailable via Microsoft API

API price $0.13/$0.52 · each benchmark row carries its own source badge (see methodology)