Model Details
Phi-3.5 MoE Instruct
Quick answer: Phi-3.5 MoE Instruct is Microsoft's August 2024 MoE instruction model scoring 91.0% ARC-C, 88.7% GSM8K, 78.5% MMLU, 59.5% MATH, and 45.3% MMLU-Pro. Apache 2.0.
At a Glance
Where Phi-3.5 MoE Instruct leads
- 91.0% ARC-C — excellent common-sense reasoning
- 88.7% GSM8K — solid grade-school math
- 78.5% MMLU — strong general knowledge
- Apache 2.0 — fully open
- MoE efficiency: high capacity at low active-param cost
Where it lags
- 59.5% MATH — moderate competition math
- 45.3% MMLU-Pro — below average for pro-level knowledge
- 37.9% Arena Hard — weaker instruction following
- August 2024: older than Phi-4 / Phi-4 Reasoning
Best for: Open Apache 2.0 MoE deployments on MMLU/ARC/GSM8K tasks; legacy Phi-3.5 ecosystem; cost-efficient knowledge retrieval.
What Phi-3.5 MoE Instruct Is
Phi-3.5 MoE Instruct is Microsoft's August 2024 mixture-of-experts instruction model in the Phi-3.5 family. It combines Phi-3's efficient training with a MoE architecture, offering competitive knowledge benchmark scores at reduced per-token compute.
It was superseded by Phi-4 (November 2024) and Phi-4 Reasoning (April 2025), both of which substantially improve on MATH and MMLU-Pro.
Specifications
| Field | Value |
|---|---|
| Organization | Microsoft |
| License | Apache 2.0 |
| HuggingFace | microsoft/Phi-3.5-MoE-instruct |
| Release date | August 20, 2024 |
| Architecture | MoE |
| Modality | Text only |
| Context window | 128K tokens |
Pricing
Open weights under Apache 2.0 — self-host at no cost. Available via Azure AI Foundry.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| ARC-C | 91.0% | Benchgen evaluation | 2024-08 |
| GSM8K | 88.7% | Benchgen evaluation | 2024-08 |
| MMLU | 78.5% | Benchgen evaluation | 2024-08 |
| MATH | 59.5% | Benchgen evaluation | 2024-08 |
| MMLU-Pro | 45.3% | Benchgen evaluation | 2024-08 |
| Arena Hard | 37.9% | Benchgen evaluation | 2024-08 |
Phi-3.5 MoE vs Alternatives
| Model | MMLU | GSM8K | MATH | License |
|---|---|---|---|---|
| Phi-3.5 MoE Instruct | 78.5% | 88.7% | 59.5% | Apache 2.0 |
| Phi-4 | — | — | — | MIT |
| Phi-4 Reasoning | — | 97.2% | 92.7% | Apache 2.0 |
| Phi-3.5 Mini Instruct | — | — | — | Apache 2.0 |
Phi-3.5 MoE vs Phi-4 Reasoning: Phi-4 Reasoning dominates on MATH (92.7% vs 59.5%) and GSM8K (97.2% vs 88.7%). Use Phi-3.5 MoE only for legacy compatibility; prefer Phi-4 Reasoning for new deployments.
Related Models & Benchmarks
- Compare: Phi-4 Reasoning · Phi-4 · Phi-3.5 Mini Instruct
- Benchmarks: ARC-C · MMLU · GSM8K
Frequently Asked Questions
What is Phi-3.5 MoE Instruct?
Microsoft's August 2024 MoE instruction model scoring 91.0% ARC-C, 88.7% GSM8K, 78.5% MMLU. Apache 2.0 — superseded by Phi-4 Reasoning for math tasks.Specs from Microsoft's Phi-3.5 MoE Instruct release (August 2024) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.