Chronologie

10 juil. 2026

Google DeepMind's Gemma 4 family offers open-weight models from 2B to 70B parameters, built for self-hosting, fine-tuning, and offline inference. Unlike ChatGPT or Claude, it's not something you chat with. It's infrastructure you control. This analysis unpacks what that means for developers, enterprises, and the open-source ecosystem.

10 juil. 2026

Cursor launched Grok 4.5, its most capable model yet, trained jointly with SpaceXAI. The real news isn't the benchmark scores: for the first time, AI agents built the training environments for the model that will eventually replace them. That self-reinforcing loop changes how fast models can improve.

10 juil. 2026

Chinese AI startup MiniMax has unleashed a wave of new models and products, including its next-generation language model MiniMax M3, an updated video generation model Hailuo 2.3, and new speech and music generation models. The company also introduced MiniMax Code, a dedicated coding assistant, signaling an aggressive expansion beyond its core conversational AI offerings.

11 juil. 2026

A consortium of researchers and industry partners released a comprehensive framework for evaluating and mitigating risks in high-stakes AI deployments. It focuses on continuous monitoring and adversarial testing, aiming to set a new baseline for safety in systems like healthcare, autonomous vehicles, and critical infrastructure. Whether developers adopt it before the next high-profile failure may determine its legacy.

11 juil. 2026

Mistral AI has raised €600 million in a funding round led by General Catalyst, the largest-ever fundraising by a French technology company. The Paris-based startup builds both open-source and proprietary large language models and intends to use the capital to hire more engineers and speed up deployment across enterprise clients in Europe and North America.

11 juil. 2026

OpenAI published an updated research roadmap this week, laying out a path to artificial general intelligence that runs through GPT-5, chain-of-thought reasoning, and real-time multimodal generation. The vision is enormous. The safety section still reads like a work in progress.

11 juil. 2026

In 2023, everyone obsessed over which model had the lowest perplexity. By 2025, that fight is over. Claude, GPT, Gemini, Qwen, they all land within striking distance on the same benchmarks. What actually separates a useful writing tool from an expensive toy is no longer the model. It's the pipeline that wraps it.

11 juil. 2026

Domyn and AISquared built their products on Ai2's fully open Olmo, Dolma, and Dolci releases, meeting auditability demands and cutting infrastructure costs by roughly half. For financial services and government clients, that transparency is the whole point.

11 juil. 2026

Moonshot AI has launched the world's first AI-native credit card with Agricultural Bank of China and American Express, tying card tiers to AI membership levels and letting users convert spending into compute tokens. The move signals a structural shift in how AI services are distributed, but raises questions about privacy and market readiness.

11 juil. 2026

OpenAI just announced a new kind of product: workspace agents, cloud-based AI workers that handle complex, multi-step jobs across teams. They run on the Codex model, live inside ChatGPT and Slack, and keep memory between sessions. It's a deliberate move from individual productivity to shared organizational intelligence. And come May 6, 2026, the free pricing flips to a credit-based model.

11 juil. 2026

A 15-part series on AI generation techniques lands less like a batch of model updates and more like an industry report. Three structural shifts emerge: multimodal pipelines have become routine, agentic orchestration is replacing manual chaining, and real-time feedback loops are turning generation from a one-shot gamble into a conversational loop.

11 juil. 2026

The M3 model team has published a detailed technical account of how they improved mathematical proof generation. The system uses layered specialist models and an evolutionary search framework called MaxProof that treats test-time computation as a guided optimization problem.

12 juil. 2026

AI coding agents can double a team's output, but only once you know how to tame them. Here's how to avoid the common traps and get consistent value from Copilot, Cursor, and similar tools.

12 juil. 2026

Most AI tools for Excel only describe how to write a formula. Kimi Sheets generates native working formulas, builds linked multi-sheet models, and outputs actual .xlsx files you can download and edit. No more copying syntax from a chatbot.

12 juil. 2026

A developer has demonstrated a full pipeline for fine-tuning Google's Gemma 4 model on Hindi text using only a free GPU, then deploying the quantized model to run inference on a standard CPU. The workflow is a blueprint for how open-weight models and cost-efficient tooling can pull low-resource language AI out of the research lab and onto devices people already own.

12 juil. 2026

Two teachers, both with the same problem: too many students, too few hours, and a stack of evidence-based practices they know would work if only there were time to implement them. Anthropic's new Claude for Teachers is the company's attempt to be the lever that closes that gap, and it comes with privacy terms that go further than most classroom AI products.

12 juil. 2026

A 26,000-student study out of China, an MIT brain-imaging experiment, and the OECD's 2026 education outlook all converge on the same finding: generative AI raises homework scores and cuts completion time, but it quietly erodes exam performance and neural engagement unless teaching is redesigned around it.

13 juil. 2026

OpenAI's latest model tops CritPt, a benchmark built from fresh, unpublished graduate-level physics problems, an approach designed to stop AI from gaming the test by memorizing answers it saw in training.

13 juil. 2026

Mistral AI launched a system of record for prompts and skills that turns scattered instructions into versioned, owned, and auditable production assets. The move addresses a growing compliance liability for enterprises running AI in customer-facing workflows, where ungoverned prompts, sitting in code repos, notebooks, or Slack threads, can embed undocumented policy decisions.

13 juil. 2026

On April 22, 2026, OpenAI launched workspace agents inside ChatGPT, turning Codex-powered assistants into persistent, shareable team members that run in the cloud, follow scheduled workflows, and integrate directly with Slack. The move widens the gap with Anthropic's Claude, whose enterprise offering remains centered on individual conversations and has no native equivalent of persistent multi-agent orchestration.

13 juil. 2026

Manus, the AI agent startup that processed 147 trillion tokens in just months, is being acquired by Meta in a deal that repositions the company as an execution layer for enterprise AI. The move signals Meta's intent to back agentic systems designed for real-world task completion, not just conversational chatbots.

13 juil. 2026

Mistral AI's model catalog in mid-2026 offers a rare window into how an AI company's strategy evolves from its experimental phase to production discipline. The portfolio, which includes 36 active models across four tiers (Featured, Frontier, Other, and Legacy), shows a company that has made hard choices about which experiments graduate and which get retired.

13 juil. 2026

Alibaba Cloud is rethinking how AI coding tools retrieve context and how cloud security defends agentic applications. A new hybrid indexing system for its Qwen-powered developer tools combines vector databases, code graphs, and pre-indexed knowledge to solve latency and privacy gaps, while its WAAP platform now wraps AI traffic with large-model-enhanced protections and autonomous security agents. These moves signal a shift from feature parity to architectural differentiation in China's cloud market.

13 juil. 2026

Alibaba's Qwen Glasses are not just another wearable. They signal a strategic bet that the post-smartphone era will be defined by the depth of AI integration into everyday vision, not by hardware specs. The tipping point for AI glasses may already be here.

14 juil. 2026

Vercel shipped Eve 0.22.6, the latest release of its public-beta framework for building AI agents whose instructions, tools, and integrations all live as plain files rather than configuration objects. The bet: inspectable files beat black-box configs.

14 juil. 2026

Perplexity swapped Grok 4.5 into Computer for Pro and Max subscribers, claiming it beat five other orchestrator configurations on an internal benchmark at roughly half the per-trial cost of Claude Opus 4.8.

14 juil. 2026

Ollama's $88 million raise leans on a number that sounds enormous and explains almost nothing: 85% of Fortune 500 companies use the platform. Cursor pulled the same move weeks earlier. These numbers aren't designed to inform. They're designed to make procurement say yes.

14 juil. 2026

Ollama, the platform that lets developers run open-weight AI models on their own hardware, just closed an $88 million round. The company is betting that open models can become the default way to run inference, and it is leaning on a playbook that Docker, whose founder Solomon Hykes now sits as an advisor, wrote two decades ago.

14 juil. 2026

Vibe coding promises to turn plain language into working software in minutes. Advocates and critics agree on one thing: the gap between prototype and production hasn't gone anywhere. As tools like Kimi Websites and Kimi Code mature, the method's trade-offs demand a closer look.

14 juil. 2026

Gartner named Cursor a Leader in its first-ever Magic Quadrant for enterprise AI coding agents. The company says 70% of Fortune 500 firms use its tools. Neither number tells you as much as it sounds like it does, and the real story is the eighteen months of product decisions that got Cursor there.

14 juil. 2026

A new AI tutoring system tested at Dartmouth College achieved an effect size between 0.71 and 1.30 standard deviations, crossing the 1.0 SD benchmark that educational interventions rarely clear. If these results hold up, the technology could reshape what personalized learning looks like at scale.

14 juil. 2026

OpenAI folded Codex directly into the ChatGPT desktop app on macOS and Windows, adding inline code editing, GitHub PR reviews without leaving the app, and support for multiple repositories in one session.

14 juil. 2026

Alibaba is betting the second half of 2026 on an integrated AI stack, not a single flagship model. Its strategy is to lock enterprises into its cloud via interoperable infrastructure, from agentic O&M to data warehousing, rather than to win benchmark titles.

14 juil. 2026

For most of the AI industry's recent history, large language models have lived inside cloud APIs and chat interfaces, processing text and images but never reaching out to touch the physical world. That boundary is dissolving faster than many analysts anticipated.

15 juil. 2026

While the industry watched OpenAI and Google, xAI quietly stitched together a multi-modal AI assistant that does more in one thread than most competitors offer across an entire product suite. Parallel multi-agent reasoning, live search, image and video generation, sub-second voice conversations, and a generous free tier: Grok is no longer just an X chatbot. It is the most complete consumer AI platform on the market, and almost nobody is talking about it.

15 juil. 2026

Cursor 2.0 ships a redesigned interface and its first fully integrated model, Composer, turning the editor from a copilot into an autonomous coding agent. The update forces a question the industry has sidestepped: should developers write code or supervise agents that write it for them?

15 juil. 2026

What if you could dictate an idea to your computer and get a spoken answer back seconds later? Here's how to build that pipeline yourself with Whisper, ChatGPT, and Python.

15 juil. 2026

15 juil. 2026

Asha Sharma told Fortune that Xbox is "not healthy" and that the company "simply spread ourselves too thin." She was not describing a bad quarter. The layoffs of 3,200 workers, the closure of four studios, and the explicit focus on only the biggest franchises paint a picture of a business being trimmed for a sale, not a turnaround.

15 juil. 2026

MiniMax has unveiled M2.7, a new AI model that excels in real-world software engineering tasks and complex office productivity workflows. The model achieves top marks on key benchmarks, including a 56.22% score on SWE-Pro, nearly matching the best Opus-level performance.

15 juil. 2026

Chinese AI lab MiniMax has released M2.5, a large language model tuned for coding and autonomous agent workflows. It claims the top spot on the Multi-SWE-Bench multilingual benchmark, and the API costs a tenth to a twentieth of comparable models.

15 juil. 2026

Chinese AI lab MiniMax has unveiled M3, a foundation model that combines state-of-the-art coding and agentic skills, a million-token context window, and native multimodality. In early benchmarks, M3 scored 83.5 on the BrowseComp agentic evaluation, beating OpenAI's Opus 4.7 at 79.3, and did it all autonomously, no humans in the loop.

15 juil. 2026

MiniMax has released Hailuo 2.3, and the most interesting thing about the update is not the motion quality or the price, it's the fact that the model now treats anime and ink wash painting as first-class outputs, not afterthoughts. That choice tells you exactly whose attention MiniMax is after.

15 juil. 2026

MiniMax has launched Speech 2.8, a new voice model that aims to make synthetic speech sound genuinely human. The update introduces native support for filler words like 'um' and 'uh', higher-fidelity timbre cloning from just 10 seconds of audio, studio-grade clarity, and improved cross-language performance starting with Chinese-to-Japanese synthesis.

15 juil. 2026

MiniMax released Speech 2.8, a voice model that builds natural filler words, hesitation, and breathing into synthetic speech. The update targets four areas: native disfluency modeling, 10-second voice cloning, studio-grade audio quality, and cross-language accuracy improvements, starting with Chinese-to-Japanese synthesis.

16 juil. 2026

Alibaba Cloud and Anthropic are pushing AI agents into the group chats where corporate work actually happens. The result looks less like a chatbot and more like a new kind of organizational model, one with roles, permissions, and asynchronous workflows.

16 juil. 2026

A new paper from Google Research proves that the creative capability of diffusion models, their ability to generate novel images and molecules beyond the training set, is not a happy accident, but a predictable consequence of neural network regularization. The finding offers a rigorous foundation for building models that balance novelty and realism.

16 juil. 2026

Nvidia's market cap surged past $3 trillion on Tuesday, and the message is hard to miss: enterprise AI infrastructure spending is still accelerating, not plateauing. Competitors are circling with cheaper chips, but the incumbent's grip on the market hasn't loosened.

16 juil. 2026

Microsoft is testing a revamped Windows 11 search menu that strips out promotional content and ads, focusing on local files, apps, and settings. The experimental version, rolling out to Windows Insiders, aims to rebuild user trust by decluttering the search experience.

16 juil. 2026

Opencode, the open-source coding agent platform, gave developers what they actually asked for: tabbed navigation. The small UX change means big things for how engineers manage multiple agent sessions and code reviews without losing their place or their context.