Vision models companies, products & suppliers
Foundation or API models whose primary modality is images or video frames. Jobs: image understanding apis; video-frame models; vision fine-tunes.
What are Vision models?
Foundation or API models whose primary modality is images or video frames. Jobs: image understanding apis; video-frame models; vision fine-tunes.
What problems does it solve?
Application teams need a vision backbone, not a consumer photo app.
Typical business use cases
- Image understanding APIs
- Video-frame models
- Vision fine-tunes
Important capabilities
- Image/video input
- Labels or embeddings
- Hosting
What buyers should evaluate
- Training-data licence
- Biometric use restrictions
- Eval set
Risks and governance considerations
A vision API used for face identification without a lawful basis.
Procurement checklist
- Intended use
- Licence
- Residency
Relevant AI Trustmark assurance
AI Trustmark independent findings appear only when an assessment or certificate exists. Category membership does not imply verification.
Products and providers
Open a product profile for its catalogue details, or use the provider link for the company profile. Claimed listings appear first, then by name. Payment never changes the order.
By Ai2
Ai2, founded by Paul Allen, conducts high-impact research and engineering to tackle key problems in artificial intelligence. Molmo2-8B is Ai2's model API, from The Allen Institute
By Apple
Depth Pro is Apple's open-source foundation model for zero-shot metric monocular depth estimation. From a single image it produces sharp, high-resolution depth maps with absolute s
By Google
SigLIP 2 So400m patch14 384 is an open-weight vision-language encoder from Google, published on Hugging Face under the Apache 2.0 licence. SigLIP 2 builds on the original SigLIP si
By IBM
Granite Vision 4.1 4B is a compact vision-language model from IBM Research, released on 29 April 2026 under the Apache 2.0 licence, built for pulling structured data out of documen
By Meta
Meta Platforms sells family-of-apps advertising products and open Llama models. Public AI pages cover research models, Llama releases and developer tooling. SAM 2.1 Hiera Large is
By Microsoft
Kosmos-2.5 is a multimodal literate model from Microsoft's Document AI research for machine reading of text-heavy images such as receipts, documents and screenshots. Pre-trained on
By NVIDIA
NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. RADIOv2.5-L is NVIDIA's mod
By Roboflow
Everything you need to build and deploy computer vision models, from automated annotation tools to high-performance deployment solutions. RF-DETR Base is Roboflow's model API, from
By Ultralytics
End-to-end computer vision platform. Annotate data, train YOLO models, and deploy to 42 global regions. Ultralytics YOLO26 is Ultralytics's model API, from Ultralytics, Inc. Publis
Products
Claimed products appear first. Ranking packs and payment do not change this list.
- Depth Pro· Apple
- Granite Vision 4.1 4B· IBM
- Kosmos-2.5· Microsoft
- Molmo2-8B· Ai2
- RADIOv2.5-L· NVIDIA
- RF-DETR Base· Roboflow
- SAM 2.1 Hiera Large· Meta
- SigLIP 2 So400m patch14 384· Google
- Ultralytics YOLO26· Ultralytics
Related categories
Relevant procurement and assurance guides
Frequently asked questions
What is Vision models?
Foundation or API models whose primary modality is images or video frames. Jobs: image understanding apis; video-frame models; vision fine-tunes.
What should not be listed as Vision models?
Products whose buyer job is finished inspection products, multimodal chat apps, or embedding-only APIs. Those belong on their own category page so search queries are not split.
Has AI Trustmark independently assessed every Vision models supplier?
No. A category listing is descriptive. Independent assessment is shown only on company or product pages that carry Trustmark evidence.
What are Vision models?
Foundation or API models whose primary modality is images or video frames. Jobs: image understanding apis; video-frame models; vision fine-tunes.
Why do buyers need to know which model providers sit behind an AI product?
Most products depend on named models or APIs. Buyers should identify those providers, where data is sent, and who is accountable when a model version changes. A supplier that cannot name material model providers is not ready for high-assurance procurement.
How can buyers tell whether customer data is used to train models?
Ask whether prompts, files, logs or outputs are used to train, fine-tune or evaluate models, including by subprocessors. Require the contractual default, any opt-out, and whether the setting can be changed silently. Treat marketing 'we do not train' claims as unverified until evidenced.
What model or provider changes should a buyer insist on being told about?
Material change usually includes a new model family, new region, new subprocessor, new write-capable tool, or a change that affects logging, privacy or human oversight. Those changes should trigger evidence refresh rather than a silent release.
How should buyers verify where AI customer data is processed?
Ask for the named processing locations, cloud regions and any subprocessors that see prompts, files or outputs. A directory listing is not evidence of residency. Independent assessment records the locations that were in scope on the assessment date.