Harrier Models by Microsoft Research

Microsoft ResearchMITOpen source

Microsoft Research releases · 6 in the last 12 months · this family litChangelog →

3 models2026Up to 32K ctx

Last refreshed 2026-05-03. Next refresh: weekly.

Details

ResearcherMicrosoft Research

LicenseMITOSI-approved

Commercial usePermitted

Models3

Released2026

Max context32K

About

Microsoft Bing's family of open-source multilingual text embedding models. Decoder-only architecture with last-token pooling. Three sizes: 270M, 0.6B, 27B. All support 32K context and 94+ languages.

Decision facts

Best fit: embeddingcoding
Capability starting point: Harrier OSS v1 27B with 32K context
Lowest tracked input: Not tracked
Closest related family: Azure Speech Services

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

3 in view

Harrier OSS v1 0.6BCurrent

Use when the workload needs embedding, 32K context, and 600M parameters.

2026-03embedding32K context600M parameters

Harrier OSS v1 270MCurrent

Use when the workload needs embedding, 32K context, and 270M parameters.

2026-03embedding32K context270M parameters

Harrier OSS v1 27BCurrent

Use when the workload needs embedding, 32K context, and 27B parameters.

2026-03embedding32K context27B parameters

Current Harrier variants with use-when guidance and lifecycle status
Model	Use when	Released	Signals	Status
Harrier OSS v1 0.6B	Use when the workload needs embedding, 32K context, and 600M parameters.	2026-03	embedding32K context600M parameters	Current
Harrier OSS v1 270M	Use when the workload needs embedding, 32K context, and 270M parameters.	2026-03	embedding32K context270M parameters	Current
Harrier OSS v1 27B	Use when the workload needs embedding, 32K context, and 27B parameters.	2026-03	embedding32K context27B parameters	Current