Best Foundation Vision Models
Compare the best 88 foundation vision models and try 76 of them on your own images, free in the Roboflow Playground. 39 are open-weight, so you can self-host them for free under their licenses.
88 models · 39 open-weight · 76 free to try · prices synced Sep 11, 2026
Open-Source Foundation Vision Models
39 models with downloadable weights you can self-host under their licenses (MIT, Apache 2.0, and Modified MIT). 27 run live in the Playground through hosted APIs, so self-hosting is optional.
| Actions | ||||
|---|---|---|---|---|
GLM 5.3 FlashZ.ai | 320B total, 18B active | MIT | Aug 2026 | |
Qwen3.8 27BQwen | 27.78B | Apache 2.0 | Aug 2026 | |
Muse Glimmer 30BMeta | 29.6B | Apache 2.0 | Aug 2026 | |
Kimi K3Moonshot AI | 2.8T | Modified MIT | Jul 2026 | |
Qwen3.6 27BQwen | 27B | Apache 2.0 | Apr 2026 | |
Qwen3.6 35B A3BQwen | 35B total, 3B active | Apache 2.0 | Apr 2026 | |
Gemma 4 26B A4BGoogle | 25.2B | Apache 2.0 | Apr 2026 | |
Gemma 4 31BGoogle | 31B | Apache 2.0 | Apr 2026 |
Foundation Vision Models via API
49 proprietary models where the weights aren't downloadable: access is through each provider's API and billed by them. Try all of them free in the Playground.
| Actions | |||||
|---|---|---|---|---|---|
GPT-6 AstraOpenAI NEW | $10.00 | $50.00 | 1.1M | Sep 2026 | |
Gemini 3.8 FlashGoogle NEW | $0.75 | $3.75 | 1.0M | Sep 2026 | |
Muse Spark 1.3Meta NEW | $1.25 | $4.25 | 1.0M | Sep 2026 | |
Claude Fable 5.1Anthropic NEW | $10.00 | $50.00 | 1M | Sep 2026 | |
Qwen3.8 FlashQwen | $0.15 | $0.47 | 1M | Aug 2026 | |
Gemini 3.7 FlashGoogle | $0.75 | $3.75 | 1.0M | Aug 2026 | |
Grok 4.6SpaceXAI | $2.00 | $6.00 | 500K | Aug 2026 | |
Muse Spark 1.2Meta | $1.25 | $4.25 | 1.0M | Aug 2026 |
How to Choose a Foundation Vision Model
Foundation models differ most in what they output, and that, not accuracy, is the first decision: language, embeddings, or masks. Pick the wrong output type and no amount of model quality saves the integration.
Need descriptions, answers, or structured text? That is the vision LLM subset. Need vectors for search, similarity, or clustering? Contrastive encoders (CLIP, SigLIP) when queries include text, self-supervised encoders (DINOv2) when precision on pure visuals matters. Need pixel-accurate regions? The SAM family. These subfamilies are not interchangeable: an embedding model cannot explain an image, and a VLM makes an expensive, imprecise segmenter.
Foundation models earn the name by adapting without full retraining, but the paths differ in cost and control. Prompting is instant and flexible (VLMs, SAM-family). Frozen features plus a small trained head is the cheap, stable middle path for classification and retrieval on encoder models. Fine-tuning buys domain accuracy at the price of training infrastructure and maintenance. Start with the cheapest path that could work; escalate only when its ceiling shows.
Foundation models frequently end up as components: SAM labels the masks that train your compact production segmenter, a VLM pre-labels the images that train your detector, CLIP embeddings bootstrap your first classifier. When you evaluate one, weigh it as a data engine and prototyping tool, not only as the thing that serves production traffic; the compact model it helps you train is often what actually ships.
The bottom line: Choose the output type first (language, vectors, or masks), then the cheapest adaptation path that could work. The foundation model that ships is often the one that trained your production model.
Frequently Asked Questions About Foundation Vision Models
Foundation vision models are large models pretrained on web-scale image or image-text data that serve as general-purpose bases rather than single-task tools. The class spans vision language models (GPT and Gemini class systems, open VLMs like Qwen VL), contrastive encoders like CLIP and SigLIP, self-supervised encoders like DINOv2, and promptable segmenters like the SAM family. What unites them is transfer: they carry enough general visual knowledge to be applied to new problems through prompting, fine-tuning, or as frozen feature extractors, instead of being trained from scratch per task. This page lists 88 foundation vision models, including 39 open-weight options you can self-host; 76 of them run live in the Playground so you can test them on your own images.
No. Language-capable VLMs are the most visible members, but CLIP, DINOv2, and SAM are foundation vision models with no language decoder: they embed, match, or segment without generating a word of text. If you need conversational answers about an image, you want the vision LLM subset; if you need embeddings, similarity, or masks at scale, the non-LLM foundation models are usually smaller, faster, and cheaper.
It depends on your task and constraints. For fixed categories in production, a model fine-tuned on your own data typically beats any general-purpose model. Compare the foundation vision models on this page and try them on your own image to see which fits.
Yes. 39 of the 88 models here are open-weight (for example GLM 5.3 Flash, Qwen3.8 27B, and Muse Glimmer 30B), free to self-host under their licenses (MIT, Apache 2.0, and Modified MIT).
Yes. You can run 76 of them in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.
This page lists all 88 foundation vision models in the Roboflow Playground catalog: 39 open-weight models you can self-host and 49 proprietary models accessed through provider APIs; 76 of them run live in the Roboflow Playground on your own images. Compare licenses, parameters, API prices, and release dates side by side, or open any model page for full details.