seminal.ai
← All open-weight models

Microsoft · General language

Phi-mini-MoE-instruct

SlimMoE compression of Phi-3.5-MoE down to 7.6B total / 2.4B active — near-Phi-3.5-MoE quality at a third of the memory, but a hard 4K context kills it for RAG. Weights only, no hosted endpoint.

Open weights · Permissive (Apache, MIT, OpenMDW) · Data checked September 6, 2026 · Not a hands-on evaluation

Model ID microsoft/Phi-mini-MoE-instruct

Parameters
7.6B total · 2.4B active7.6B-A2.4B MoE
Architecture
Mixture of experts
Context window
4,096 tokensMaximum input
Max output
Not recordedPer response
Input price
Not published
Output price
Not published
Cached input
Not published
Blended price
Not available3:1 input to output
Knowledge cutoff
2023-10
Released
2025-06-23
Status
ga
Size band
4B to 15B

License and openness

Phi-mini-MoE-instruct is released under the MIT. This is a permissive license. It generally allows use, modification and redistribution, including commercial use, subject to attribution and notice requirements. License terms can change between versions; confirm the text published with the weights you download.

Capabilities

The reference records no capability tags for this model.

Input modalitiestext
Output modalitiestext

What it is for

Weights only, no hosted endpoint.

Cost at list price

InputNot published
OutputNot published
Cached inputNot published
Blended, 3:1Not available

The provider publishes no per-token price for this model. Any price you see elsewhere belongs to a third-party host, and running the weights yourself has hardware costs instead.

Estimate a workload across all models ↗

Sources

Sources checked September 6, 2026. Specifications and prices change without notice; confirm against the provider before you commit.

Other open-weight models from Microsoft

See all 23 from Microsoft ↗

Browse all open-weight models ↗