DEVELOPER REFERENCE · VERIFIED MODEL EVIDENCE

Microsoft
Compare 22 source-verified model records, repository licenses, published weight formats and dated evidence. No overall model ranking.
Evidence statusSource-first verified
Latest record check2026-11-13
Latest record full verification2026-11-01
DIRECT ANSWER · SOURCE-OBSERVED REGISTRY
What Microsoft model evidence is available?
Open Model Weights currently tracks 22 field-verified Microsoft model records. In this indexed sample, the recorded modalities are 10 Text; 3 Image; 3 Image + text → text; 2 Image To 3D; 1 Audio → text; 1 Not declared; 1 Screenshots and text → GUI/terminal actions; 1 Text → audio. The published repository license labels include 19 mit; 2 apache-2.0; 1 cc-by-nc-sa-4.0. The listed weight formats include Safetensors (17), PyTorch (11), ONNX (1).
Coverage describes the model repositories indexed here, not every model released by the publisher. A checked source, an unknown field and an unchanged source revision are different states. No missing value is treated as a negative answer.
22verified model records
8recorded modalities
3distinct license labels
22model repository links
Microsoft Research and other Microsoft teams publish downloadable AI checkpoints for language, vision, document processing, speech and 3D generation. Its verified repository entries range from the Phi small-language-model family to Florence, TRELLIS, VibeVoice and encoder-based models.
These are source-checked model repositories, not a complete Microsoft Foundry catalog. Hosted Microsoft services, specialized decision-scoring APIs and Azure deployment products are separate from downloadable weights; the existence of a Phi checkpoint does not make other Microsoft models open weight.
01Phi language models versus Florence vision models
The Phi series covers compact language and selected vision-language checkpoints, whereas Florence-2 focuses on visual tasks conditioned by text prompts, including image understanding and captioning. Phi-4 and Phi-3.5-vision-instruct therefore need different prompts, processors and evaluations; a single parameter count is not a capability ranking.
02TRELLIS image-to-3D research checkpoints
Microsoft's TRELLIS and TRELLIS.2 research projects publish image-to-3D pipelines and downloadable artifacts. These models generate 3D representations or assets, not ordinary text completions. Compare supported input images, artifact size, GPU/runtime dependencies, output export and the specific release repository.
03VibeVoice and document-specific tasks
VibeVoice-1.5B is described by Microsoft's project as a long-form text-to-speech system, while VibeVoice-ASR is a separate transcription checkpoint. LayoutLMv3 and Table Transformer address layout and document/table understanding. These are distinct model families with different license requirements and inference input pipelines.
04Repository licenses are model-specific
Most checked Microsoft records show MIT labels, but the registry also includes Apache-2.0 records and a CC BY-NC-SA 4.0 label for LayoutLMv3-base. That label is not a generic Microsoft license; it signals that commercial reuse should not be assumed for that checkpoint. Check the precise revision and license text for any redistribution.
05Open weights are separate from commercial endpoints
Microsoft also sells model access through cloud services. A hosted Foundry capability is not evidence that a corresponding model checkpoint is redistributable. For an OMW model record, require an independently verified artifact repository and model-specific license before describing it as downloadable open weights.
Publisher background is attributed to official sources. Model counts and license classifications come from OMW repository evidence; no model benchmark or processing location is inferred.
Textmit
Language and instruction reasoning
Inspect Phi-4 for its exact architecture, weights and MIT license; confirm runtime and prompts before use.
Image + text → textmit
Visual question answering
Inspect the Phi-3.5 Vision repository for its vision processor and multimodal inputs.
Image + text → textmit
Vision task prompting
Use Florence-2-base as the starting point for visual understanding and image-text outputs.
Image To 3Dmit
Image-to-3D generation
Review TRELLIS.2-4B for 3D pipeline, artifact exports and system dependencies.
Text → audiomit
Text-to-speech generation
Examine VibeVoice-1.5B's documented multi-speaker speech generation workflow.
Not declaredcc-by-nc-sa-4.0
Document layout processing
Inspect LayoutLMv3's non-commercial CC BY-NC-SA 4.0 label rather than assuming universal MIT permission.
Compare models →Explore weight-memory estimates →Understand licensing →
FIELD-BY-FIELD COMPARISON · PRIMARY LINKS
Model evidence at a glance.
Compare the exact entries in our registry. Parameter values are taken from the source-checked record; a license label is not a legal opinion, and recognized weight files are not an inference-speed benchmark. Open each record for the underlying citations and observed history.
Compare two to four models →Explore weight memory →Browse observed history →
LICENSE & COVERAGE · NOT LEGAL ADVICE
How do the recorded licenses differ?
Different models from the same developer can use different license terms. The distribution below describes the labels and field-level classifications observed in the listed repositories; it does not automatically establish permissions for a specific deployment.
License labels
- mit19 of 22
- apache-2.02 of 22
- cc-by-nc-sa-4.01 of 22
Commercial-use classifications
- allowed21 of 22
- not-allowed1 of 22
Model modalities
- Text10 of 22
- Image3 of 22
- Image + text → text3 of 22
- Image To 3D2 of 22
- Audio → text1 of 22
- Not declared1 of 22
- Screenshots and text → GUI/terminal actions1 of 22
- Text → audio1 of 22
EVIDENCE METHOD · TRANSPARENT LIMITS
How this developer profile is assembled.
We aggregate the current source-checked model records, rather than publish an unverified corporate biography. Repository metadata, model cards, license files and recognized weight artifacts supply the facts where available. The record-check date does not assert that a model or its license changed on that date.
Source-derived runtime signals are not hardware benchmarks. Weight-only memory estimates omit runtime overhead, activations and other deployment requirements. Model-specific uncertainties remain visible in the individual evidence records.
Evidence policy →Full methodology →Independent measurements →
QUESTIONS · DIRECT ANSWERS
Frequently asked questions about Microsoft models.
Which Microsoft models are covered?
Open Model Weights currently lists 22 field-verified records for Microsoft. Their recorded model modalities are: 10 Text; 3 Image; 3 Image + text → text; 2 Image To 3D; 1 Audio → text; 1 Not declared; 1 Screenshots and text → GUI/terminal actions; 1 Text → audio. This is a registry sample, not an exhaustive publisher catalog.
Which licenses are recorded for Microsoft models?
In the 22 indexed records, repository license labels are: 19 mit; 2 apache-2.0; 1 cc-by-nc-sa-4.0. License labels do not replace the model-specific terms or establish rights for every use case.
Where can I verify Microsoft model information?
Each indexed model links to its official source repository and a field-level Open Model Weights profile. The most recent record checks shown here reach 2026-11-13; checking a repository does not imply that its contents changed.
Are Microsoft's Phi and Florence models used for the same task?
No. The published Phi collection covers language and multimodal instruction models, while Florence-2 is designed for prompted image-understanding tasks. The model cards identify their distinct inputs and libraries.
Is every Microsoft model in this registry MIT-licensed?
No. The checked entries include MIT and Apache-2.0 labels, but LayoutLMv3-base carries a CC BY-NC-SA 4.0 label. A model's rights cannot be transferred from another Microsoft release.
Does availability in Microsoft Foundry prove downloadable model weights?
No. An API or managed cloud offering is a service, not a reproducible checkpoint. OMW lists independently checked weight repositories and records the separate licensing evidence for each.
COMPLETE INDEXED SELECTION
All 22 indexed model profiles.
Florence-2-base
Image + text → text
1,024 tokens
mit
Field-verified
512 tokens
mit
Field-verified
TRELLIS-image-large
Image To 3D
Not declared in checked config/model card
mit
Field-verified
Not declared in checked config/model card
mit
Field-verified
512 tokens
mit
Field-verified
BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext
Text
512 tokens
mit
Field-verified
table-transformer-structure-recognition
Image
1,024 tokens
mit
Field-verified
512 tokens
mit
Field-verified
VibeVoice-ASR
Audio → text
131,072 tokens
mit
Field-verified
layoutlmv3-base
Not declared
514 tokens
cc-by-nc-sa-4.0
Field-verified
512 tokens
mit
Field-verified
Florence-2-large
Image + text → text
4,096 tokens
mit
Field-verified
Phi-3.5-mini-instruct
Text
131,072 tokens
mit
Field-verified
131,072 tokens
mit
Field-verified
Phi-3-mini-4k-instruct
Text
4,096 tokens
mit
Field-verified
Phi-3.5-vision-instruct
Image + text → text
131,072 tokens
mit
Field-verified
table-transformer-detection
Image
1,024 tokens
mit
Field-verified
16,384 tokens
mit
Field-verified
VibeVoice-1.5B
Text → audio
65,536 tokens
mit
Field-verified
Not declared in checked config/model card
apache-2.0
Field-verified
2,048 tokens
mit
Field-verified
Mage-VL-Cowork 4B
Screenshots and text → GUI/terminal actions
262,144 tokens
apache-2.0
Field-verified