AI Daily: Cursor and Kimi Model Controversy, Claude's New Features, and Mistral's Content Tax

Behind the Cursor Code Myth: The Unexpected Appearance of Kimi and Recent AI Highlights

Did you know that sometimes the most eye-catching news in the tech world comes from an accidental discovery? In today’s rapidly evolving environment, every update to development tools affects countless engineers. Recently, an interesting topic emerged on social media that has sparked intense discussion among developers, even drawing attention to AI regulatory trends as far away as Europe.

Daily tech news is always full of surprises. Let’s dive into the major events that have recently impacted how we work.

Is Cursor Truly “Exposed” This Time?

It all started when a user named Fynn was testing the popular AI code editor Cursor. While playing around with OpenAI’s base URLs, they unexpectedly captured an unvarnished string of code: “accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast”.

This URL was like a secret that couldn’t be kept. It turns out that the highly acclaimed Composer 2 model, which even received public support from Elon Musk, is actually built on the foundation of Kimi K2.5. After this discovery was retweeted by Musk and viewed over 730,000 times, it immediately went viral. Some commenters joked that this completely stripped away the mystery of a high-valuation company, with one user even quipping, “at least change the model ID.”

Facing public curiosity, the team did not dodge the issue. Aman Sanger, a member of the Cursor team, later confirmed the news. He explained that the team conducted rigorous evaluations based on perplexity across many base models and ultimately found Kimi k2.5 to be the strongest performer.

To put it simply, it’s like buying a top-tier sports car chassis and then using your own expertise to perform precision engine tuning. The Cursor team performed continuous pre-training (CPT) and high-compute reinforcement learning (RL) on top of Kimi, expanding the overall compute scale by four times. Combined with a powerful underlying architecture and Fireworks’ inference sampler, they successfully pushed Composer-2 to the forefront of technology. Aman admitted that failing to mention Kimi in the official release was a PR oversight and promised to improve in the future.

However, behind this seemingly settled “PR oversight” lies a deeper commercial licensing controversy. In fact, the “Moonshot AI” team behind Kimi wasn’t so generous at first. Du Yulun, their pre-training lead, posted that the tokenizers of both models were identical and questioned why Cursor’s founder didn’t follow the license agreement or pay any fees. Other employees also posted similar questions, though these posts were later quietly deleted.

This involves Kimi K2.5’s open-source license terms: if a commercial product’s monthly revenue exceeds $20 million, it must prominently display the “Kimi K2.5” label on the product interface. Cursor’s current annual recurring revenue (ARR) has reached $2 billion (approx. $167 million per month), which is more than eight times the threshold.

Although both parties eventually reached a graceful settlement—with Cursor admitting the “lack of labeling was an oversight” and Kimi officially “retweeting congratulations”—this incident undoubtedly punctured the “self-developed myth” of some high-valuation AI applications in Silicon Valley. A rising star aiming for a $50 billion valuation has its core model weapon borrowed from Kimi, which is valued at $4.3 billion. This highlights the delicate dependency between base models and applications and makes one reconsider whether such technical moats are as impregnable as they seem.

The Rival’s Counterattack: Windsurf Offers a Seven-Day Gift Pack

Business reaction speed is always impressive. Since Kimi K2.5’s strength has been publicly validated by a top team, other competitors weren’t going to miss this marketing opportunity.

The well-known AI editor Windsurf quickly announced on social media that seeing the renewed public interest in Kimi K2.5, they decided to offer a limited-time promotion. For the next seven days, all users across Trial, Pro, Teams, and Max tiers can experience the relevant features for free.

This is a clever marketing tactic. By leveraging the buzz created by a competitor to promote their own platform, they allow developers curious about Kimi to test it out without friction. This also highlights the intense competition in the AI code editor market, where every small technical advantage or social topic can become a key trigger for user migration.

How Claude Cowork Reshapes Project Management

Beyond the intense competition in the coding field, everyday office applications are also seeing exciting upgrades. If you find yourself dealing with mountains of documents and fragmented tasks daily, Claude’s newly launched Cowork platform project feature is worth exploring.

This new feature focuses on centralizing all tasks and contexts within a single workspace. Previously, workers often had to switch frequently between different folders and chat windows. Now, all files and instructions can be kept securely on your local machine, significantly reducing privacy concerns. Users can easily import existing projects with a single click or start entirely new workflows from scratch. The AI assistant has evolved from a passive Q&A tool into a virtual project manager that is truly integrated into daily operations and possesses high organizational capabilities. By delegating tedious data organization to automation tools, humans can reserve valuable time for creative decision-making.

Europe’s AI Dilemma: Can Mistral CEO’s “Content Tax” Proposal Solve It?

While tech giants compete with flashy new features, the tug-of-war between regulation and copyright remains a harsh reality for the industry. Recently, Mistral AI CEO Arthur Mensch wrote a thought-provoking piece for the Financial Times, detailing Europe’s struggles in the competition and potential solutions.

The European continent has nurtured a wealth of cultural and creative ideas, which are invaluable soft power assets. However, major companies in the US and China currently train models using vast amounts of public content in a very loose environment. In contrast, European developers are trapped in a restrictive legal framework. The current “opt-out” mechanism is difficult to implement in practice, as copyrighted works continue to circulate online while protection mechanisms remain fragmented and ineffective.

Facing a dual deadlock where copyright holders worry about their livelihoods and AI developers face legal uncertainty, Mistral proposed a new concept: a “revenue-based levy” (content tax) on all providers launching commercial AI models in the European market.

The key point is that this tax would apply equally to overseas providers. This means foreign companies operating in Europe must contribute for the public content they use. These funds would flow into a central fund dedicated to investing in new content creation and supporting the cultural industry. In return, development teams would receive the legal certainty they desperately need, exempting them from infringement liability for training on public data. Mensch emphasized that creators and developers are not enemies but natural allies. This proposal has sparked widespread discussion and invites reflection on whether a clever balance can be found between protecting creators and maintaining technical competitiveness.

From Cursor’s model secrets to Claude’s practical project management upgrades and Mistral’s macro reflections on industry regulations, the pace of the tech industry never slows. These advancements are not just technical iterations but real-life examples of how humans and digital tools are shaping future lifestyles together.

Q&A

Q1: Did Cursor violate open-source license terms by using the Kimi model? A: There is a strong suspicion of violation based on the terms. According to Kimi K2.5’s license, if a commercial product’s monthly revenue exceeds $20 million, it must prominently display “Kimi K2.5” on the interface. Currently, Cursor’s ARR is $2 billion (approx. $167M/month), which is over eight times the threshold. While both parties eventually settled with Cursor admitting it was a “PR oversight,” many speculate this was an emergency fix after being caught.

Q2: Is it common in the AI world to fine-tune someone else’s model and package it as “self-developed”? A: It seems to be becoming more common. Besides Cursor’s Composer 2, Windsurf has previously admitted to using China’s Zhipu GLM model. Additionally, Cognition (developers of Devin) had its SWE-1.5 model confirmed via prompt injection to be based on GLM-4.6. This practice of “using an open-source model as a base, adding a layer of fine-tuning, and packaging it as self-developed without attribution” has become a common pattern for some AI companies to save on high training costs.

Q3: Will Cursor’s future market position be threatened as giants enter the field? A: It is facing significant challenges. Giants with their own base models are changing the game. For example, Anthropic’s Claude Code captured a high market share in just 8 months, earning a 46% “favorite” rating in a recent developer survey (compared to Cursor’s 19%). Meanwhile, OpenAI’s Codex desktop app surpassed one million downloads in its first week. These giants with self-developed models don’t rely on external bases, giving them massive cost advantages and pricing power, which poses a serious threat to companies like Cursor whose “moat” is built on someone else’s model.

Q4: What protections does Claude’s new Cowork project feature offer for privacy-conscious workers? A: One of Claude Cowork’s biggest selling points is “local security.” The company explicitly states that when using the project feature, “Files and instructions stay on your computer.” This means users can centralize tasks and contexts and import projects with one click while significantly reducing the risk of leaking sensitive business data.

Q5: Is the “content tax” proposed by the Mistral CEO only for European AI companies? A: No, the key to this proposal is “equal treatment.” The tax would apply to all providers of commercial AI models in the European market, including overseas tech giants from the US and China. Any company operating in Europe would pay this revenue-based fee, which would go into a European central fund to support cultural and content creation industries. In exchange, both local and overseas AI developers would gain “legal certainty,” exempting them from infringement liability for using public data for training.