AI Labs Landscape

Founded: 2015 (as nonprofit), restructured 2019 (capped-profit). Headquarters: San Francisco.

Key models: GPT-2 (2019), GPT-3 (2020), GPT-4 (2023), GPT-4o (2024), o1 (2024), o3 (2025), GPT-5 (2025), GPT-5.1 (Nov 2025), GPT-5.3-Codex (Feb 2026), GPT-5.3 Instant (Mar 2026), GPT-5.4 (Mar 2026), GPT-5.5 and GPT-5.5 Pro (Apr 23, 2026; API Apr 24, 2026).

Technical contributions:

• Scaling laws for neural language models (Kaplan et al., 2020)
• In-context learning via large-scale pretraining (GPT-3)
• RLHF pipeline for alignment (InstructGPT, Ouyang et al., 2022)
• Reasoning via reinforcement learning (o-series models)
• Multimodal input/output (GPT-4o: text, image, audio, video)

Research focus (2025-2026): Reasoning models, adaptive test-time compute, agentic coding, computer use, tool search, multimodal perception, and safety evaluation. The 2026 GPT-5.x line shows the specialist and general tracks moving in parallel: GPT-5.3-Codex targets coding agents, while GPT-5.4 is the broader reasoning and tool-use model.

Business model: API access, ChatGPT subscription products. Closed weights for frontier models.

Founded: 2021 by Dario Amodei, Daniela Amodei, and others (ex-OpenAI). Headquarters: San Francisco.

Key models: Claude 1 (2023), Claude 2 (2023), Claude 3 family (2024: Haiku, Sonnet, Opus), Claude 3.5 Sonnet (June 2024), Claude 3.7 Sonnet (Feb 2025, introducing extended thinking), Claude 4 family (May 2025), Claude 4.5 family and Haiku 4.5 (late 2025), Claude 4.6 family (Feb 2026), Claude 4.7 Opus (April 2026).

Technical contributions:

• Constitutional AI (Bai et al., 2022): alignment via written principles rather than purely human preference labels
• Mechanistic interpretability: circuit analysis, dictionary learning, sparse autoencoders for understanding model internals
• Scaling monosemanticity (Templeton et al., 2024): extracting interpretable features from large models
• Extended thinking: visible chain-of-thought reasoning in Claude 3.5+

Research focus (2025-2026): Interpretability at scale, AI safety evaluations, responsible scaling policies, alignment science. Anthropic publishes unusually detailed interpretability and system-card work for a closed frontier lab.

Business model: API access, Claude consumer product. Closed weights.

Founded: DeepMind (2010, acquired by Google 2014), Google Brain (2011). Merged into Google DeepMind (2023). Headquarters: London and Mountain View.

Key models: AlphaGo (2016), AlphaFold (2020, 2024), PaLM (2022), Gemini 1.0 (2023), Gemini 1.5 (2024), Gemini 2.0 (late 2024), Gemini 2.5 (2025), Gemini 3 (Nov 2025), Gemini 3.1 Pro and Flash variants (2026), Gemma open models through Gemma 4 (Apr 2026), and on-device Gemini Nano / Gemma variants.

Technical contributions:

• AlphaFold: solved protein structure prediction (Nobel Prize, 2024)
• Transformer (Vaswani et al., 2017, at Google Brain)
• Natively multimodal training (Gemini: text, image, audio, video from scratch)
• Long context (Gemini 1.5: 1M+ token context window)
• Mixture-of-Experts architectures at scale
• Gemma: open-weight models for research and on-device deployment

Research focus (2025-2026): Multimodal reasoning, science applications (materials, genomics), long context, audio/voice models, efficient open models, and agentic research workflows. Google DeepMind has one of the broadest public research portfolios among frontier labs, spanning pure science (AlphaFold) and consumer products (Gemini in Google services).

Business model: Integrated into Google products (Search, Workspace, Cloud). Gemini API. Gemma models released open-weight.

Founded: 2023 by Elon Musk. Headquarters: Austin, Texas.

Key models: Grok-1 (2023, 314B MoE, released open-weight), Grok-2 (2024), Grok-3 (2025).

Technical contributions: Large-scale MoE training. Grok-1 was one of the first frontier-scale models released with open weights. Grok-3 trained on one of the largest GPU clusters assembled (100K+ H100s).

Research focus (2025-2026): Scaling compute, real-time information integration (via X/Twitter data), reasoning capabilities.

Business model: Integrated into X platform, API access.

Founded: Facebook AI Research (FAIR) in 2013. Headquarters: Menlo Park.

Key models: Llama 1 (2023), Llama 2 (2023), Llama 3 (2024), Llama 3.1 (2024, up to 405B), Llama 4 family (2025): Scout, Maverick, and Behemoth, introducing Meta's first frontier MoE models.

Technical contributions:

• Llama family: established the open-weight model ecosystem
• Self-supervised vision: DINO, DINOv2 (self-distillation with no labels)
• JEPA (Joint Embedding Predictive Architecture): Yann LeCun's proposed alternative to autoregressive and generative approaches
• Segment Anything (SAM): foundation model for image segmentation
• No Language Left Behind (NLLB): multilingual translation

Research focus (2025-2026): Open-weight frontier models, self-supervised learning, world models (V-JEPA), embodied AI, multilingual coverage. Meta's strategy is to release open-weight models that become ecosystem standards, driving adoption of Meta's infrastructure.

Business model: Models released under permissive licenses. Revenue from Meta platforms (ads), not model APIs.

Founded: 2023. Headquarters: Paris, France.

Key models: Mistral 7B (2023), Mixtral 8x7B (2023, MoE), Mistral Large (2024), Mistral Large 2 (July 2024, 123B), Mistral Nemo (2024, 12B, joint release with NVIDIA), Ministral 3B and 8B (Oct 2024, edge-optimized), Codestral (code), Codestral Mamba (state-space variant), Mistral Small and Medium (2024-2025 commercial tiers).

Technical contributions:

• Efficient open-weight models: Mistral 7B matched Llama 2 13B performance
• MoE at accessible scale: Mixtral 8x7B demonstrated strong MoE results with modest hardware requirements
• Sliding window attention for efficient long-context processing
• Ministral line: sub-10B models targeting edge and on-device inference

Research focus (2025-2026): Efficient architectures, multilingual European models, enterprise deployment. Mistral occupies a strategic position as the leading European AI lab, with implications for EU AI regulation and data sovereignty.

Business model: Open-weight base models, commercial API, enterprise licenses for larger models.

Founded: 2023 as an AI subsidiary of High-Flyer (Chinese quantitative hedge fund). Headquarters: Hangzhou, China.

Key models: DeepSeek-Coder (2023), DeepSeek-V2 (2024, 236B MoE), DeepSeek-V3 (Dec 2024, 671B MoE with 37B active), DeepSeek-R1 (Jan 2025), DeepSeek-R1-0528 (May 2025), DeepSeek-V3.1 (Aug 2025), DeepSeek-V3.2-Exp (Sep 2025), DeepSeek-V3.2 and V3.2-Speciale (Dec 2025). As of April 21, 2026, DeepSeek-R2 has not been documented in DeepSeek's official release notes, so this page does not treat R2 as a shipped model.

Technical contributions:

• Highly efficient MoE training at frontier scale
• Auxiliary-loss-free load balancing for MoE routing
• Multi-head Latent Attention (MLA) for KV-cache compression
• DeepSeek-R1: demonstrated that long-chain reasoning can emerge from RL on verifiable math and code tasks; R1-Zero used pure RL, while the public R1 model added cold-start and post-training stages for readability and instruction following
• DeepSeek-V3.1 / V3.2: hybrid thinking and non-thinking modes, stronger tool use, agent tasks, and sparse-attention experiments for long context
• Open release of full model weights and detailed technical reports, narrowing the gap between closed frontier labs and open releases

Research focus (2025-2026): Efficient pretraining, RL-driven reasoning, open-weight frontier models, sparse attention, FP8 training, and tool-using agent tasks. DeepSeek has had outsized impact relative to compute budget by publishing detailed engineering decisions.

Business model: Open-weight models plus a commercial API with aggressively low pricing.

Founded: Qwen team within Alibaba Cloud, first public release 2023. Headquarters: Hangzhou, China.

Key models: Qwen (2023), Qwen2 and Qwen2.5 (2024), Qwen3 family (2025: dense models 0.6B-32B, Qwen3-30B-A3B, Qwen3-235B-A22B, and API-only trillion-parameter Qwen3-Max variants).

Technical contributions:

• Among the strongest open-weight Chinese-English bilingual model families through 2025-2026
• Qwen2.5-Math and Qwen2.5-Coder: specialized reasoning-and-code variants that competed with closed models on math and programming benchmarks
• Qwen3: hybrid thinking / non-thinking modes and open weights across a wide parameter range, with 100+ language and dialect coverage

Research focus (2025-2026): Open-weight scaling, multilingual coverage, specialized reasoning and code variants, long-context extensions.

Business model: Permissive open-weight releases plus Alibaba Cloud API integration.

Founded: 2023. Headquarters: Beijing, China.

Key models: Kimi (2023, long-context chat), Kimi k1.5 (2025, multimodal reasoning), Kimi K2 (July 2025, open-weight MoE), Kimi K2.5 (Jan 2026), Kimi K2.6 (Apr 2026, open-weight multimodal MoE, 1T total parameters with 32B active and 256K context).

Technical contributions:

• Kimi K2 is among the largest open-weight MoE releases from a Chinese lab; weights publicly available on Hugging Face
• Deployed Muon optimizer at scale for LLM pretraining; Moonshot published "Muon is Scalable for LLM Training" (arXiv:2502.16982, 2025), one of the first large-scale validations of the Muon update rule outside toy settings
• Long-context specialization: Kimi models targeted 128K+ context windows early in their development cycle

Research focus (2025-2026): Frontier open-weight MoE, efficient optimization (Muon), long-context coding, multimodal agents. The Muon training paper is directly relevant to optimizer research — it provides empirical scaling evidence that Nesterov-momentum-based spectral updates transfer from small-scale experiments to production LLM training.

Business model: API and consumer chat (kimi.ai), open-weight releases for research.

Founded: 2019 by Aidan Gomez (Transformer co-author), Ivan Zhang, Nick Frosst. Headquarters: Toronto.

Key models: Command family (Command, Command R, Command R+).

Technical contributions:

• Retrieval-Augmented Generation (RAG) as a first-class product feature
• Enterprise-focused model design: grounded generation with citations
• Multilingual embeddings (Cohere Embed)

Research focus (2025-2026): Enterprise NLP, grounded generation with source attribution, domain-specific fine-tuning, efficient deployment.

Business model: Enterprise API, on-premise deployment options.

Founded: 2022. Headquarters: San Francisco.

Key focus: Open-source AI infrastructure. Provides APIs for running open-weight models (Llama, Mistral, etc.) at scale. Contributes to open model training and fine-tuning tooling.

Technical contributions:

• FlashAttention integration and optimized inference infrastructure
• Open model training runs (RedPajama dataset, Together model series)
• Efficient fine-tuning APIs for open models

Research focus (2025-2026): Inference optimization, open model ecosystem, and making capable open models easier to deploy.

Business model: Cloud API for open models, enterprise infrastructure.

Founded: 2016. Headquarters: New York and Paris.

Key focus: The central hub for the open ML ecosystem. Hosts models, datasets, and demo applications (Spaces).

Technical contributions:

• Transformers library: the standard Python interface for pretrained models
• Hub: hosts 500K+ models, 100K+ datasets as of 2025
• Tokenizers, Datasets, Accelerate, PEFT, TRL libraries
• Democratized access to model weights, training, and evaluation

Role in 2025-2026: Hugging Face is not a model developer in the frontier sense. It is infrastructure. When Llama, Mistral, or DeepSeek release weights, they land on Hugging Face. When researchers share fine-tunes, adapters, or datasets, they go to the Hub. Understanding the Hugging Face ecosystem is a practical requirement for working with open models.

Founded: 2024 by Ilya Sutskever (ex-OpenAI Chief Scientist), Daniel Gross, and Daniel Levy. Headquarters: Palo Alto and Tel Aviv.

Key focus: Building safe superintelligence as a single, focused goal. No products, no API, no revenue pressure. Pure research toward superintelligent AI with safety guarantees.

Technical contributions: Not yet public as of April 21, 2026. The significance of SSI is primarily about who is involved (Sutskever co-authored the sequence-to-sequence paper, co-led GPT-2/3/4 development, and was OpenAI's Chief Scientist) and the organizational thesis: that safety and capability research must be unified from the start, not bolted on after.

Research focus: Alignment, scalable oversight, and capability research toward superintelligence. Details are not public.

Founded: late 2025 by David Silver (UCL professor; ex-DeepMind reinforcement-learning team lead for over a decade; principal investigator on AlphaGo, AlphaZero, MuZero, and AlphaProof). Headquarters: UK (specific city not disclosed).

Funding: Announced a 1.1 billion USD seed round at a 5.1 billion USD post-money valuation on April 27, 2026 — co-led by Sequoia Capital and Lightspeed Venture Partners, with Nvidia, Google, DST Global, Index Ventures, and the UK Sovereign AI Fund participating. Multiple outlets reported it as the largest seed round in European history.

Stated mission: Build a "superlearner" that acquires knowledge from its own experience via reinforcement learning, rather than from human-generated data. The intellectual scaffold is the Silver–Sutton essay Welcome to the Era of Experience (2025; forthcoming in MIT Press volume edited by George Konidaris), which argues that scaling on human-generated data is approaching its limits and the next regime is agents that generate their own training signal through interaction.

Technical contributions: None yet. No published research under the Ineffable Intelligence name as of April 28, 2026; no disclosed product, benchmark, or specific capability claim. Continuity with Silver's prior work is the only visible technical thread; the open research question is whether verifier-grounded RL (the AlphaZero / AlphaProof family) extends to open-ended domains where the reward signal is not unambiguous.

Why it is on this page: It is the single largest seed-stage AI funding event in Europe to date, and the technical thesis (RL beats human-data scaling at the frontier) is a substantive bet against the prevailing pretraining + RLHF paradigm. See the dedicated Ineffable Intelligence page for the detailed treatment, the "what is not yet known" list, and the references.

Founded: 2022 by researchers from DeepMind, Google Brain, FAIR. Headquarters: San Francisco.

Key models: Reka Edge (2024, 7B, on-device), Reka Flash (2024, 21B, fast API), Reka Core (2024, frontier multimodal).

Technical contributions:

• Multimodal from the ground up: Reka Core processes text, images, video, and audio natively, without separate vision encoders bolted onto a language backbone
• Competitive multimodal benchmarks with smaller parameter counts than frontier labs' equivalents

Research focus (2025-2026): Efficient multimodal architectures, on-device deployment (Reka Edge). Primarily an API and enterprise play rather than an open-weight contributor.

Business model: API access, enterprise licensing.

Founded: 2017. Headquarters: Tel Aviv, Israel.

Key models: Jurassic series (2021-2023), Jamba (2024, hybrid SSM-Transformer), Jamba 1.5 (2024, 52B active / 398B total).

Technical contributions:

• Jamba: one of the first production hybrid architectures combining Mamba (SSM) layers with Transformer attention layers; the hybrid approach targets the memory efficiency of SSMs for long contexts while retaining attention's in-context recall for short-range dependencies
• Demonstrated that SSM-Transformer hybrids can reach competitive quality at moderate scale with lower KV-cache memory than pure-Transformer equivalents

Research focus (2025-2026): Hybrid SSM-Transformer architectures, long-context efficiency, enterprise NLP.

Business model: API access, enterprise products (Wordtune, Writer tools).