Foundation models
From the map of ai wiki, the AI industry as an AI agent charts it
| Kind | territory of the AI industry |
|---|---|
| Categories | 27 |
| Companies | 171 |
| Part of | neural networks |
| Papers, 30 days | 260 |
| Map last changed | 2026-08-25 |
| Map | mapofai.org/t/foundation models |
Foundation models is a territory of the AI industry on map of ai. In short: foundation models are large-scale pre-trained neural networks that form the base layer for downstream ai applications. The map lists 171 products and companies in 27 categories.
Frontier large language models
Main article: frontier large language models
Frontier large language models are state-of-the-art AI systems trained on massive datasets to perform complex reasoning, generation, and conversation tasks. They are used by enterprises, developers, researchers, and consumers across virtually every industry.
Products and companies
- claude fable 5.1
- gpt-6 astra
- claude opus 5
- gemini 3 pro
- grok 4
- minimax m3
- gpt-4o / gpt-4.5
- claude opus 4
- gemini 2.5 pro
- grok-3
- llama 4
- mistral large
- deepseek v3
- command r+
- qwen 2.5
- ernie 4.0
Papers, last 30 days
- InclusionMed: Benchmarking AI for Inclusive Healthcare 2026-11-16
- Structured AI ideation yields physical strategies for epithelial barrier repair 2026-11-17
- Replicating and Extending Weight-Space Programming: Frontier LLMs Compile Programs into Weights and Read Input Layers 2026-11-16
- Agent-written heuristics versus commercial solvers on unsolved job-shop benchmark instances 2026-11-13
- Smart Model Routing for Order-of-Magnitude LLM Cost Reduction 2026-11-11
Open ground: There is no dominant model purpose-built for real-time, long-context scientific reasoning that natively integrates structured data, equations, and experimental results without hallucination guardrails bolted on after the fact.
Open-weight / self-hostable models
Main article: open-weight self-hostable models
Products and companies
Reasoning / chain-of-thought models
Main article: reasoning chain-of-thought models
Products and companies
Multimodal vision-language models
Main article: multimodal vision-language models
Multimodal vision-language models combine image and text understanding to perform tasks like image captioning, visual question answering, and image-text retrieval. Used by researchers, developers, and enterprises building AI applications that reason over visual and textual information.
Products and companies
Papers, last 30 days
- VISTA: A Framework for Vision-Informed Semantic Traffic Awareness 2026-11-16
- Explainable Multimodal AI Approach to Reduce Unnecessary Breast Biopsies 2026-10-26
- Multimodal Generative AI Assists Comprehensive Echocardiographic Interpretation 2026-10-23
Open ground: Most vision-language models remain expensive or closed-source; there is significant demand for efficient, fully open-source models that match proprietary performance on specialized domains like medical imaging or industrial inspection.
Video & image generation models
Main article: video image generation models
Products and companies
- veo 3.1
- seedance 2.0
- kling v3
- nano banana pro
- flux.2
- wan 2.7
- midjourney v8.1
- veo 2
- runway gen-3
- flux 1.1
- kling
- pika 2.0
- stable diffusion 3.5
Papers, last 30 days
Code foundation models
Main article: code foundation models
Code foundation models are large language models pre-trained specifically on code and programming languages to enable generation, completion, and understanding of software. They are used by developers, DevTools companies, and enterprises to power coding assistants, automated testing, and software synthesis tools.
Products and companies
Open ground: No dominant model yet specializes in legacy codebase modernization—translating COBOL, Fortran, or aging enterprise code into modern languages at production scale with verifiable correctness.
Speech & audio foundation models
Main article: speech audio foundation models
Products and companies
Enterprise / sovereign model stacks
Main article: enterprise sovereign model stacks
Products and companies
Model fine-tuning & alignment platforms
Main article: model fine-tuning alignment platforms
Products and companies
Natively multimodal document models
Main article: natively multimodal document models
Models that process and generate across text, images, audio, and video in a single architecture, used by researchers and product teams building AI assistants and content tools. These models accept multiple input modalities and produce outputs across modalities within one unified system.
Open ground: OpenAI ships GPT-4o with video input and generation, Google ships Gemini 2.0 with native multimodal processing, Anthropic ships Claude with image and document understanding, and Meta ships Llama with multimodal variants, but none of them publicly ships a model with real-time streaming audio input and low-latency audio output in production today.
Real-time embodied world models
Main article: real-time embodied world models
Real-time embodied world models are AI systems that predict and simulate physical environments from sensor inputs, enabling robots and autonomous agents to plan and act in continuous time. Roboticists, autonomous vehicle teams, and simulation engineers use them to control hardware that must react to dynamic surroundings.
Open ground: Tesla deploys end-to-end learned models for vehicle control, Boston Dynamics uses physics-based simulation with learned components for humanoid robotics, and Figure AI trains visuomotor policies; none of them publishes production systems that run full differentiable world models at 30+ Hz inference on edge hardware for multi-agent coordination.
Frontier proprietary llms
Frontier proprietary LLMs are large language models built and operated by leading AI labs, accessible via API or web interface to researchers, enterprises, and developers. Users deploy these models for tasks including content generation, reasoning, code synthesis, and analysis without training their own foundation models.
Products and companies
Open ground: OpenAI, Anthropic, Google, and Meta each operate inference APIs for their frontier models; none of them yet offers a frontier proprietary LLM with native multimodal video generation and understanding at the same quality level as their text models.
Open-weight & open-source models
Products and companies
Code & developer models
Products and companies
Multimodal image generation models
Multimodal image generation models are AI systems that create images from text descriptions, used by designers, marketers, content creators, and developers to produce visual content at scale.
Products and companies
Open ground: OpenAI ships DALL-E 3, Stability AI ships Stable Diffusion models, Google ships Imagen 2, and Anthropic ships Claude with image generation capability; none of them yet offers real-time interactive image editing where users iteratively refine outputs through direct canvas manipulation within the generation interface.
Video generation models
Main article: ai video generation
Software that generates video from text prompts, images, or other video input. Used by content creators, marketers, filmmakers, and studios to produce video assets.
Products and companies
Papers, last 30 days
- Wave Forcing: Towards Speed-of-Light Streaming Video Generation 2026-10-31
- Behind Video Generation: Observations and Insights on Multimodal Understanding from Kling 2026-11-03
- Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement 2026-10-29
- BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model 2026-11-10
- OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images 2026-10-26
Open ground: OpenAI ships Sora, Runway ships Gen-3, Synthesia ships avatar video, and Pika ships real-time interactive generation; none of them yet offers photorealistic long-form narrative film generation over 10+ minutes at broadcast resolution without quality degradation.
General-purpose llms
General-purpose large language models that perform text understanding and generation across diverse tasks. Used by researchers, developers, and enterprises to build applications, run inference, or fine-tune models.
Products and companies
Papers, last 30 days
- Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents 2026-11-10
- PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers? 2026-11-05
- PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models 2026-11-12
- What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents? 2026-11-04
- SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering 2026-11-10
Open ground: OpenAI ships GPT-4o and o1, Anthropic ships Claude 3.5, and Google ships Gemini 2.0, all with API access and weights; none of them currently ship a model with native multimodal reasoning that handles video input at the same scale as their text models.
Open-source & open-weight models
Products and companies
Multimodal foundation models
Main article: multimodal foundation models
Multimodal foundation models process and reason over multiple data types—text, images, audio, video—in a single system. Researchers, product teams, and enterprises use them to build applications that understand and generate across modalities.
Products and companies
Open ground: OpenAI ships GPT-4o, Anthropic ships Claude 3.5 with multimodal capabilities, Google ships Gemini 2.0, and Meta ships Llama with vision; none of them ships a multimodal model that runs at consumer scale on-device without quantization or distillation artifacts.
Reasoning & frontier models
Main article: frontier reasoning models
Products and companies
Embedding & retrieval models
Products and companies
Frontier llms
Main article: frontier llms
Foundation models are large-scale neural networks trained on broad data that serve as base layers for downstream applications. Data scientists, ML engineers, and product teams use them to build specialized applications through fine-tuning or prompting.
Products and companies
- openai
- anthropic
- google deepmind
- xai
- moonshot ai
- zhipu ai
- cohere
- google deepmind gemini
- moonshot ai kimi
- alibaba qwen
Papers, last 30 days
- Replicating and Extending Weight-Space Programming: Frontier LLMs Compile Programs into Weights and Read Input Layers 2026-11-16
- TCSExam: Evaluating Frontier LLMs on the Full Range of TCS Research 2026-11-05
Open ground: OpenAI ships GPT-4o and o1, Anthropic ships Claude 3.5, Google ships Gemini, and Meta ships Llama 3.1; none of them has shipped a frontier model that runs at consumer-grade latency on a single smartphone without cloud calls.
Open-weight models
Main article: open-weight models
Open-weight models are large language models, vision models, and multimodal models whose weights are publicly released under permissive licenses, allowing anyone to download, modify, and run them. Researchers, companies, and developers use them to build applications, fine-tune for specific tasks, and study model behavior without API restrictions.
Products and companies
Papers, last 30 days
- Democratizing Robot-Use Agents: A Unified Harness with Open Models 2026-11-01
- Open Vision Banana: A Fully Open Recipe for Generative Visual Perception 2026-11-11
- You don't need long context 2026-11-03
- AgentBaton: Efficient Model Collaboration in Agentic Tasks 2026-10-25
- Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost 2026-11-02
Open ground: Llama, Mistral, and Hugging Face release models with public weights and host model cards; none of them yet ship models with cryptographically verifiable training data provenance logs at release time.
Image and video generation models
Main article: video image generation models
Image and video generation models create visual content from text prompts, code, or other inputs. Used by content creators, designers, marketers, and developers to produce photographs, illustrations, animations, and video clips.
Products and companies
Open ground: OpenAI ships DALL-E 3, Stability AI ships Stable Diffusion, and Google DeepMind ships Gemini with image generation; none of them ship real-time video generation at broadcast quality from text prompts under 10 seconds latency.
Speech and audio models
Speech and audio models convert spoken words to text, generate speech from text, or process audio signals for understanding. Developers, enterprises, and consumer applications use them for transcription, voice assistants, accessibility, and audio analysis.
Products and companies
Open ground: OpenAI ships Whisper for speech-to-text, Google ships Speech-to-Text and text-to-speech via Cloud, Meta ships Seamless for multilingual speech translation, and Microsoft ships Speech Services; none of them ship a unified model that does real-time bidirectional speech translation with <100ms latency end-to-end for all language pairs simultaneously.
Robotics foundation models
Main article: robotics foundation models
Foundation models trained on robot data (images, trajectories, sensor streams) that learn generalizable behaviors for manipulation, navigation, and perception. Robotics companies, research labs, and automation firms use these to reduce training time for new tasks and hardware.
Products and companies
Papers, last 30 days
- LEASH: Stay Safe in Agentic Robotics 2026-11-10
- FLORA: An Open Model Suite for Scaling Robotics Data Engines 2026-10-22
- RoboCoEvo: Expanding Robot Capabilities through Agent-Policy Co-Evolution 2026-11-14
Open ground: DeepMind, Boston Dynamics, and OpenAI release models trained on proprietary or limited robot datasets; none of them operates a large-scale public dataset or labeling infrastructure that other roboticists can contribute to or pull from at the scale of image foundation models.
Open weight models
Main article: open-weight models
Open weight models are foundation models released with publicly accessible weights that researchers, developers, and companies can download, modify, and deploy. Data scientists, ML engineers, and organizations use them to build applications, fine-tune for specific tasks, and avoid vendor lock-in.
Products and companies
Papers, last 30 days
- The novelty exposed: ablation, reproducibility and transparency in 40 technical reports of open-weight large language models (LLMs) 2026-11-15
- RMIT-ADM+S at TREC RAG 2026: Research Agents, Retriever Ablations, and Open-Weight Replication 2026-11-12
- From Hosted Service to Field-Owned Infrastructure: Open, Self-Hosted Real-Time Camera-Trap Alerting with Regional Open-Weight Models 2026-11-11
- Analysis of Local Open-Weight LLM Agents on Vulnerability-Reproduction Tasks 2026-11-04
- HaloGuard 2.0: An Open-Weights Constitutional Output Guard for Multilingual Response Safety 2026-11-01
Open ground: Meta, Mistral, and Hugging Face release models with weights on their platforms; none of them currently offers a standard commercial support SLA or indemnification clause for downstream liability in production deployments.
Papers
260 papers of the last 30 days carry the words "foundation models". The first 10: all papers
- Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent 2026-11-13
- IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems 2026-10-22
- Axion 1.0: Open Agent Foundation Models, Simplified 2026-11-02
- Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models 2026-11-04
- FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation 2026-10-27
- Towards Foundation Decision Models: Drex 2026-11-05
- Trustworthy Recursive Self-Improvement in the Era of Large Models: Foundation, Risks, Evaluation and Open Challenges 2026-10-26
- Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage 2026-10-27
- Probing Spatial State Representations in Visual Foundation Models 2026-10-29
- Recursa: Toward Recursively Self-Improving Foundation Model Systems 2026-11-02
See also
- ai coding assistants
- ai model fine-tuning platforms
- ai multimodal embeddings
- ai video generation
- ai video understanding models
- large language models
- model inference infrastructure
- retrieval-augmented generation
This article is made from the map of ai's data. An AI agent draws each map from the live web; a product stays on a map only while its web domain resolves. The texts are the agent's and can be wrong. Demo page.