Microsoft
Efficient

Phi-3.5 MoE

MoE

MoE variant of Phi-3.5, 16x3.8B experts (~6.6B active). Better quality at similar cost to dense Phi-3.5 mini.

Phi-3.5 MoE is a efficient AI model from Microsoft. It costs $0.1600 per million input tokens and $0.6400 per million output tokens (blended $0.3040/M), with a 128,000-token context window.

RatesUSD per million tokens. Blend is 70% input, 30% output, our calculation.

INPUT
$0.1600/M
per million input tokens
OUTPUT
$0.6400/M
per million output tokens
BLENDED 70/30
$0.3040/M
unchanged since 3 May
CONTEXT
128,000
tokens

Hand-maintained from Microsoft's published pricing; no check date on record.

What it is good at
  • MoE efficiency
  • Open weights
  • 128K context
Typical use cases
  • Self-hosted small chat
  • MoE research

Benchmarks

vs. best public score
MMLU78.9%
Multitask academic knowledge across 57 subjects. Source: Microsoft (self-reported), MMLU, read 2026-11-10.
MATH59.5%
High-school competition math problems. Source: Microsoft (self-reported), MATH, read 2026-11-10.
HumanEval70.7%
Python function synthesis from docstrings. Source: Microsoft (self-reported), HumanEval (0-shot), read 2026-11-10.
Hand-curated from each provider's published reports and public leaderboards. Methodology varies across sources, treat as directional rather than authoritative.

More from Microsoft

See all 8 →

Frequently asked questions

How much does Phi-3.5 MoE cost?

Phi-3.5 MoE costs $0.1600 per million input tokens and $0.6400 per million output tokens, for a blended reference rate of $0.3040 per million tokens.

What is Phi-3.5 MoE's context window?

Phi-3.5 MoE supports up to 128,000 tokens of context in a single request.

What is Phi-3.5 MoE best for?

Phi-3.5 MoE is well suited to MoE efficiency, Open weights and 128K context.

Who makes Phi-3.5 MoE?

Phi-3.5 MoE is developed and served by Microsoft. It was released in Aug 2024.

Terms used on this page