Startup labs

Kai-Fu Lee's lab behind the open Yi models, now focused on enterprise AI products.

Humanoid robotics company, founded as Halodi Robotics, building home robots and a world model for robot learning.

Agent startup that trained models to act in software, releasing the open Persimmon-8B and Fuyu-8B multimodal models.

Jamba, the first production-scale hybrid of Transformer and Mamba layers.

German company building sovereign LLMs for enterprises and governments, including Pharia models and tokenizer-free T-FREE research.

R&D lab from Jeremy Howard and Eric Ries turning foundational research into practical tools, such as FSDP+QLoRA and ModernBERT.

The Claude models, Constitutional AI and interpretability research on what happens inside models.

US company releasing open-weight models such as AFM and Trinity, and creator of the MergeKit model-merging toolkit.

Beijing AI company, one of China's AI Tigers, known for the open Baichuan language models and medical models.

The FLUX image models, from the researchers behind latent diffusion and Stable Diffusion.

Real-time voice AI company pioneering new model architectures, maker of the Sonic text-to-speech models.

Maker of wafer-scale AI chips whose research team released open Cerebras-GPT and BTLM language models.

Company building AI models of biochemical structure and interaction for molecular design, starting with Chai-1.

Consumer chatbot company founded by LaMDA developers Noam Shazeer and Daniel de Freitas, training its own conversational models.

Data-centric AI company built on confident learning, a method for finding label errors in datasets.

San Francisco company behind Devin, an autonomous AI software engineer, and its own SWE agent models.

Enterprise models (Command, Embed, Rerank) and Cohere Labs' open multilingual Aya research.

Enterprise RAG company founded by RAG co-author Douwe Kiela, known for RAG 2.0 and the KTO alignment method.

American AI and robotics company, formerly Embodied Intelligence, headquartered in Emeryville, California.

Data curation company that researches how selecting and synthesizing training data makes models better and cheaper.

Real-time generative video lab, known for Oasis, a playable AI-generated game world model.

San Francisco lab releasing open Cogito models trained with iterated distillation and amplification.

German AI research company known for the DeepL Translator and its own translation-focused language models.

Frontier-class open-weight models trained cheaply: DeepSeek-V3's efficient MoE and R1's open reasoning recipe.

The company of DuckDB's creators, the in-process analytics database born at CWI; Amazon agreed to buy it in 2026.

AI audio research and product company known for natural-sounding speech synthesis and voice agent models.

San Francisco lab led by Ashish Vaswani, publishing open pretraining research, datasets and the Rnj-1 model.

Frontier AI lab for the life sciences, maker of the ESM3 generative protein language model and ESM Cambrian.

Humanoid robot company founded by Brett Adcock, developer of the Figure robots and the Helix vision-language-action model.

Therapeutics company pioneering generative biology, known for the Chroma generative model for protein design.

Video generation lab that released Mochi 1, an open-source text-to-video model.

Builds low-latency voice AI models for conversational agents; the $100M total includes a $30M extension of its seed round.

Paris startup building computer-use agents and the open Holo models that navigate user interfaces.

Company building mathematical superintelligence; its Aristotle system reached gold-medal level at the 2025 IMO.

The Hub for open models and datasets and the Transformers library; also BLOOM, StarCoder, FineWeb and SmolLM.

Voice AI company grounded in emotion science, known for the Empathic Voice Interface and Octave text-to-speech.

Maker of the Ideogram text-to-image models, noted for rendering legible text inside generated images.

Chinese speech and language AI company, known for voice recognition and, with HIT, the Chinese BERT-wwm and MacBERT models.

AI lab founded in 2021 that trained a 70B model from scratch and builds tools that put software under users' control.

AI research and product company building diffusion-based large language models, starting with Mercury.

Public benefit corporation that built Pi, a personal AI assistant powered by the Inflection AI model stack.

Search AI company founded by Han Xiao, known for open embedding, reranker and reader models.

Frontier AI lab building generative models of biology, known for Latent-X for de novo protein binder design.

French enterprise generative AI company, co-developer of ModernBERT and maker of late-interaction retrieval models.

Jerusalem creative-app company behind Facetune and the open LTX-Video real-time video generation models.

Liquid foundation models (LFM) for devices, built on liquid neural network research from MIT CSAIL.

Generative video company behind Dream Machine and the Ray video models, with research on pretraining such as Inductive Moment Matching.

AGI startup building code models with ultra-long context, including 100M-token context windows and the HashHop eval.

Beijing computer vision company behind Face++, with research on efficient networks such as ShuffleNet and the YOLOX detector.

Self-described independent research lab led by David Holz, known for the Midjourney text-to-image models.

Lightning attention for million-token context in open models, plus Hailuo video and speech models.

Europe's leading model maker: efficient open-weight models, Mixtral's mixture of experts, and Le Chat.

Tsinghua-incubated Beijing company best known for the open MiniCPM small language and on-device multimodal models.

The Kimi models: long context early, then the open-weight Kimi K2 trained with the Muon optimizer.

Company behind GPT4All for local LLMs and the reproducible Nomic Embed text embedding models.

The open Hermes fine-tunes, YaRN context extension, and decentralized training over the internet.

The GPT models and ChatGPT, and the scaling results that set off the large language model race.

AI research company building AI scientists and autonomous laboratories for the physical sciences.

AI answer engine company that also releases open models such as the pplx-embed embedding family.

General-purpose robot foundation models: the pi-zero vision-language-action family, open-sourced as openpi.

AI company developing foundation models for software development, including the open-weight Laguna coding models.

Japanese deep learning company that created Chainer and Optuna and develops the PLaMo Japanese language models and its own AI chips.

San Francisco company training models over decentralized compute, known for the INTELLECT series and RL environments.

AI protein design company that created OpenCRISPR-1, described as the first AI-designed gene editor.

AI drug discovery company that builds foundation models of cell biology from large-scale microscopy data.

Lab founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou to build open frontier models.

Natively multimodal models (Core, Flash, Edge) from former DeepMind and Meta researchers.

New York generative video company behind the Gen-3 and Gen-4 video models and the GWM-1 world model.

Ilya Sutskever's lab with one goal and one product, a safe superintelligence; it has published nothing yet.

Nature-inspired methods: evolutionary model merging and The AI Scientist, from a Transformer co-author.

Bengaluru company building language models for Indian languages, part of India's sovereign LLM programme.

Scale AI, Inc. is an American artificial intelligence infrastructure and software company based in San Francisco, California.

Hong Kong-listed AI company known for computer vision research and the SenseNova foundation model series.

Conversational voice company whose Conversational Speech Model aims to cross the uncanny valley of AI voice.

Robotics company building an omni-bodied robot foundation model, the Skild Brain, to control any robot for any task.

Data-centric AI company spun out of the Stanford AI Lab, originating from the Snorkel weak supervision project.

Stable Diffusion, the open image model release that opened up generative imaging in 2022.

Shanghai AI company, one of China's AI Tigers, releasing open speech, video and multimodal reasoning models.

Music generation company whose models create songs with vocals from prompts; earlier released the open Bark audio model.

London AI video company that researches neural rendering and synthetic avatars for business video.

Mira Murati's lab: Tinker, a fine-tuning API, and open research notes on reproducible inference.

AI cloud company co-founded by researchers including Tri Dao, with research such as RedPajama and Mixture-of-Agents.

Video-native AI company with offices in San Francisco and Seoul, building the Marengo and Pegasus video models.

Open-source toolkit maker for fast, memory-efficient fine-tuning and local running of LLMs, with many quantized models.

South Korean AI company known for the Solar language models and depth up-scaling, plus document AI.

Embedding and reranking model company for retrieval, acquired by MongoDB in 2025.

Self-driving company founded by Raquel Urtasun, using generative AI simulation such as UniSim for autonomous trucks.

London autonomous driving company using end-to-end learning, known for the GAIA generative driving world models.

Fei-Fei Li's spatial intelligence company: models that generate explorable 3D worlds (Marble).

Enterprise generative AI platform company that trains its own Palmyra family of large language models.

The GLM models, spun out of Tsinghua University's KEG lab; GLM-4.5 and later are open weights.

San Francisco lab training open multimodal models, known for Zamba SSM-hybrid language models and Zonos text-to-speech.