Players

The companies, labs and models shaping the field

20 articles · 3 sous-catégories

Amazon Nova — AWS's model strategy facing OpenAI and Google

Launched in December 2024, the Nova family is Amazon's first genuinely competitive line of proprietary AI models. Aggressive pricing, native Bedrock integration, Nova 2 with extended reasoning: a balanced overview of an actor playing its own card in a landscape dominated by OpenAI, Anthropic and Google.

amazonnovabedrockawscanvasreelnova-2multimodalhyperscalertitanpricingagentic

profil Companies

Cohere — enterprise-first AI from Toronto

Co-founded by a co-author of the Transformer paper, Cohere builds LLMs exclusively for enterprises: private deployments, native RAG, extended multilingualism. A portrait of a B2B player that refuses to compete with its own customers.

coherecommand-rcommand-aayacohere-labsrerankembedenterpriseragmultilingualismb2btorontoai-sovereignty

profil Companies

DeepSeek — when a quantitative fund builds a frontier laboratory

DeepSeek AI is a Chinese laboratory founded in 2023 by a quantitative hedge fund. It releases its models as open weights and has produced documented architectural advances, notably with V3 and R1.

deepseekr1v3moeopen-weightsreasoninggrpomlahigh-flyerliang-wenfengdistillation

profil Ecosystem

Hugging Face — the platform that distributes open-source AI to the world

Hugging Face is not a frontier research lab: it is the distribution infrastructure that makes models accessible. Hub, libraries, Enterprise, Open-R1 — anatomy of an ecosystem that has become indispensable.

huggingfacetransformershubdatasetsspacesopen-r1gradioargillaenterpriseopen-sourceml-libraries

profil Models

Qwen — Alibaba Cloud's open-weight model family

Tongyi Qianwen (Qwen) is the large language model family developed by Alibaba Cloud since 2023. By 2026 it had become the top contributor of derivative models on HuggingFace, combining an exceptional range (0.6B to 235B), hybrid dense/MoE architecture, and integrated reasoning.

qwenalibabatongyi-qianwendamotongyi-labopen-weightsmixture-of-expertsreasoningmultimodalapache-2-0modelscopechina

profil Companies

xAI and Grok — the AI of real-time data

xAI is Elon Musk's AI lab, founded in 2023 to develop Grok, an AI assistant integrated with X. A factual profile: history, models, Colossus infrastructure, and safety controversies.

xaigrokelon-muskfrontiercolossusllmreal-timemixture-of-expertsopen-source

profil Companies

Anthropic — from research to product

Founded by former OpenAI executives, Anthropic bets on AI safety to build Claude, one of the most advanced models on the market.

anthropicclaudeai-safetyfrontier-labsfunding

profil Models

The Claude family — Haiku, Sonnet, Opus, and now Fable

Anthropic's Claude lineup held to three tiers from 2024. It has had four since the summer of 2026, and the most capable one is no longer called Opus.

claudeopussonnethaikumodelsanthropic

concept Companies

Constitutional AI — when the model self-aligns against written principles

Constitutional AI is an alignment method where a model critiques and revises its own outputs against a list of principles, without human annotation of harmful outputs.

constitutional-airlaifanthropicalignmentrlhf

concept Ecosystem

The LLM Ecosystem — mapping the forces at play

Who builds the large language models, who controls the infrastructure, and how geopolitical competition is reshaping the landscape. A structural reading of the ecosystem in 2025-2026.

ecosystemfrontier-labsopen-sourcestrategycompetition

profil Models

Gemini — Google's answer

Multimodal by design, Gemini is the model family with which Google DeepMind catches up with and challenges the competition.

geminigooglemultimodalmodels

profil Companies

Google DeepMind — infrastructure in service of AI

Born from the merger of DeepMind and Google Brain, Google DeepMind combines cutting-edge fundamental research with unmatched compute infrastructure.

googledeepmindinfrastructureresearch

profil Models

OpenAI's GPT models — from GPT-4o to Astra

The GPT family has defined the LLM market. From GPT-4o to the reasoning models, then to GPT-5's router — and to the return of explicit choice with the 5.6 lineup.

gptopenaio1o3reasoningmodels

profil Models

Llama — Meta's take on open source

With Llama, Meta released an open-weights model family that transformed the open-source LLM ecosystem.

llamametaopen-weightsmodels

profil Companies

Meta AI — the open source bet

Meta chose to release its Llama models as open weights, a strategic decision that reshuffled the cards in the AI ecosystem.

metaopen-sourcellamastrategy

profil Companies

Mistral AI — the European alternative

Founded in 2023 by French researchers from Meta and Google, Mistral AI became Europe's fastest AI unicorn. A clear-eyed look at a genuine achievement and its structural tensions.

mistraleuropefrancelabs

profil Models

Mistral models — from Mixtral to Medium 3.5

From the open-source Mistral 7B to the proprietary Mistral Large, a European model lineup built around efficiency and Mixture of Experts architecture.

mistralmixture-of-expertsopen-sourcemodels

comparatif Ecosystem

Open source vs proprietary — what it actually means for LLMs

A model labelled 'open source' is not always what it claims to be. A breakdown of the distinctions that matter: licences, data, performance and regulation.

open-sourceproprietaryllamamistralstrategylicence

profil Companies

OpenAI — from non-profit to commercial leader

Founded as a non-profit organization, OpenAI has become the most highly valued AI company in the world with ChatGPT and the GPT series.

openaichatgptgptfrontier-labs

concept Companies

Google TPU — The only credible alternative to Nvidia GPUs

Since 2016, Google has been deploying its own AI chips in production. A radically different architecture from GPUs, built for scale — but difficult to use outside its ecosystem.

tpugooglehardwareinfrastructuretraining