☕️ Alibaba's First 1T Model, Mistral's $1.7B Series C & A Painful $1.5B Settlement for Anthropic

From Europe to the USA, investor money continues to flow into AI startups from last week's craze, while Alibaba's Qwen debuts its first 1T model.

Here are our main key takeaways:

Join us at AI Tangle as we untangle this week's happenings in AI!

​​​Alibaba's AI team Qwen debuts its biggest AI model to date with over 1 trillion parameters

​Announced in an X/Twitter post earlier this week, Alibaba's Qwen team has announced its biggest AI model to date, with over 1 trillion parameters as it puts its hopes on scale. Dubbed Qwen3-Max-Preview (Instruct), the non-reasoning preview model reportedly beats out many of its rivals, such as DeepSeek-V3.1, Kimi K2, and Claude Opus 4, though leaves out any from xAI, OpenAI, or Google. Starting at $0.86 per million tokens for shorter prompts, the rather pricey model is available on Qwen's chatbot website and OpenRouter.

​

Mistral AI raises €1.7 billion in funding as ASML becomes its leading shareholder

Mistral AI's recently rumored fundraiser has been locked in and finalized as the French startup nets itself €1.7 billion in Series C funding. Dutch chip equipment manufacturer ASML joined for the first time and led heftily, contributing €1.3 billion of the total and gaining an 11% stake in Mistral. The €11.7 billion post-money valuation round more than doubles Mistral's previous valuation of €5.8 billion in June of last year. With so much more capital behind it, Mistral states it will continue to build the next chapter of AI, investing it in infrastructure and expansion.

​

Anthropic agrees to a $1.5 billion payout to settle a class action lawsuit by authors

​Anthropic will pay $1.5 billion to settle a class-action lawsuit from authors who alleged that the company used pirated books as training data for its Claude chatbot. Called the largest copyright recovery in history, the settlement additionally requires Anthropic to destroy all downloaded copies of pirated content; however, it does not include an admission of liability. The settlement fund amounts to $3,000 for 500,000 downloaded books, but it could grow if more works are identified, setting a precedent that could influence similar suits against OpenAI, Microsoft, and Meta.

​

OpenAI blends its Model Behavior research team into a larger one amid reshuffling

OpenAI is reorganizing its Model Behavior team, the ones responsible for shaping ChatGPT's personality and reducing bias, into its larger Post Training group as model behavior becomes central to core development. The team's leader, Joanne Jang, is starting a new in-house unit called OAI Labs to explore novel AI-human interaction interfaces. The restructuring comes not long after online criticism of GPT-5's colder tone and ongoing scrutiny over chatbot behavior, including a lawsuit that alleges that ChatGPT contributed to a teenager taking their own life.

​

Microsoft and Nebius sign a 5-year AI compute deal worth up to potentially $20 billion

To sate its ever-growing thirst for more power, Microsoft has signed a 5-year deal worth up to $20 billion with AI infrastructure group Nebius to lease GPU compute. Nebius, a spin-off of Russia's Yandex, says the contract's baseline value is $17.4 billion, but it could rise to $19.4 billion if Microsoft's demands increase. The deal will finance the construction of Nebius' new data center in Vineland, New Jersey, boosting its U.S presence. Meanwhile, the company's shares skyrocketed by 68% following the news, while its rival, CoreWeave, rose 5%.

​

Cognition AI bags $400 million in fresh capital at a valuation of $10.2 billion in a recent fundraiser

Cognition AI has raised $400 million at a $10.2 billion valuation, up from $4 billion earlier this year, Bloomberg reports. Led by Founders Fund with extra backing from the likes of Lux Capital and 8VC, the round comes after a massive surge in demand for its coding agent Devin and many others like it, which boosted Cognition's ARR to $73 million in June from just $1 million last September. Cognition had also acquired fellow AI coding startup Windsurf in July, while at the same time laying off ~30 employees as workers face brutal 80-hour workweeks.

​​​​​Uxia - Uxia makes user testing accessible for design and product teams of all sizes by leveraging AI and synthetic users to deliver fast, reliable insights and feedback.

​Gem - Unify your recruiting tech stack with Gem, a platform powered with AI, CRM, and analytics to provide an all-in-one solution to recruiters' problems.

​HATCH - Combine natural interaction with a chatbot and an infinite canvas for co-creation in Hatch, a collaborative canvas powered by AI tools that work together and work with you.

​Averi - Design and build your marketing strategy in an all-in-one AI-powered platform to create content and build a perfect-fit team that intros to human marketeers when you need them.

​​​​​Why Do AI Language Models Hallucinate? (5-min read)

Hallucinations in AI powered by large language models has remained as one of its core problems since inception, but is there more to it? A new research paper by a team at OpenAI certainly alleges so, explaining how current AI training affects hallucination and why benchmarks need to change.

📻 AI Confidential Podcast - Agents Are the New API Client​

💼 The Artificially Intelligent Enterprise - The Human Side of AI​

💡 AI CIO - The Shrinking Window of Defense​

🎯 The AI Marketing Advantage - Inside OpenAI's New Plan to Expand Opportunity With AI​