Chinese social media giant RedNote, also known by its Chinese name Xiaohongshu, has released its first large language model as open source, deepening Beijing’s push to expand its influence in artificial intelligence despite rising geopolitical barriers.

The model, called dots.llm1, was made available on developer platform Hugging Face on Friday, accompanied by a technical paper outlining its capabilities and architecture.

The move aligns RedNote with other major Chinese tech firms like Alibaba and DeepSeek that have recently adopted open-source strategies in AI. Unlike U.S. giants OpenAI and Google, which have kept their most advanced models proprietary, China’s AI developers have turned to open source not just to build credibility and developer ecosystems but also to challenge Western dominance in the rapidly evolving sector.

RedNote’s dots.llm1 is a mixture-of-experts model containing 142 billion parameters, but it only activates 14 billion parameters per query. The design prioritizes computational efficiency without significantly compromising performance. According to the company’s paper, the model was trained on over 11.2 trillion tokens of real-world data—intentionally excluding synthetic sources to improve trustworthiness and accuracy.

Benchmark tests show dots.llm1 performs competitively against Alibaba’s Qwen 2.5 series in programming tasks, while it falls short of top-tier models like DeepSeek-V3. On C-SimpleQA, a Chinese language benchmark, dots.llm1 scored 56.7—slightly below DeepSeek-V3’s 68.9, but close to Qwen2.5-72B in multilingual and mathematical evaluations.

RedNote’s research team also published intermediate training checkpoints for every trillion tokens, allowing the academic community to track the model’s learning curve over time. This transparency further distinguishes the project in a space where many companies still guard technical details.

RedNote’s open-source move comes at a critical time for China’s AI sector. With the United States maintaining strict export controls on advanced semiconductors and high-performance GPUs, Chinese firms are looking for alternative routes to assert technological leadership.

These companies are building global developer networks, strengthening diplomatic ties through technological cooperation, and asserting China’s relevance in AI policy discourse, by releasing models like dots.llm1 to the public.

The open-source strategy has proven effective for others in China’s tech ecosystem. Earlier this year, Alibaba launched Qwen 3, the latest in its series of public AI models, while startup DeepSeek released its low-cost R1 model, earning global attention for achieving strong performance at a fraction of the cost of Western counterparts.

DeepSeek’s newer V3 series has been ranked in benchmarks alongside OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet. With dots.llm1 now in the open-source arena, RedNote is clearly aiming to join this elite group of Chinese innovators challenging U.S. leadership in AI.

RedNote’s Global Expansion

Best known as an “Instagram-like” social app popular among China’s Gen Z, RedNote gained wider global attention earlier this year when some American users began migrating to the platform amid fears of a TikTok ban in the U.S.

The company has since opened an overseas office in Hong Kong and expanded its product offerings. Its AI strategy includes Diandian, a new AI-powered search assistant that helps users navigate Xiaohongshu content—an indicator that the firm is preparing for more global and AI-integrated products.

Since 2023, shortly after OpenAI launched ChatGPT, RedNote has stepped up investment in large language models. Dots.llm1 is its most ambitious public AI initiative to date.

The release of dots.llm1 has added to a growing perception that China is moving quickly—and openly—to compete in AI development, despite export bans and geopolitical tension. With more Chinese firms embracing open source, the global AI industry is becoming more distributed and competitive.

While American firms like Meta have also pursued open-source models, many others have opted for closed systems, citing safety and competitive concerns. RedNote’s entry into open-source AI deepens the divide between the American and Chinese approaches to AI deployment, governance, and access.

But it may also force global policymakers and tech leaders to reckon with a more multipolar future for artificial intelligence—one where China, with or without U.S. semiconductors, plays a central role in shaping how machines learn and reason.

As developers begin experimenting with dots.llm1 and creating finetuned derivatives, RedNote’s strategy could further shift the momentum in favor of open access, and in doing so, extend China’s soft power in digital infrastructure.

After pouring over $13.5 billion into OpenAI, Microsoft is finally addressing growing concerns around the profitability and long-term sustainability of the high-stakes alliance. For the first time, CEO Satya Nadella has spoken publicly about the evolving nature of the partnership—acknowledging changes that hint at underlying business tensions, even as both sides maintain they are committed to working together.

The remarks come at a time when questions are mounting about whether Microsoft’s massive investment is translating into direct financial returns, especially as OpenAI pursues its own ambitious expansion plans, such as the $500 billion Stargate project and new infrastructure deals with rival cloud providers.

In an interview with The Circuit’s Emily Chang (Bloomberg), Nadella admitted the collaboration is no longer what it once was. “Any company that has gone from being a research lab to one of the most successful product companies of this age — obviously things have to change for them and for us and in the context of the partnership,” he said, emphasizing that while the relationship is “evolving,” it remains strong.

Microsoft’s Profit Puzzle

At the core of the shift lies a fundamental issue: profitability. While Microsoft has become a leader in enterprise AI tools—embedding OpenAI’s models in everything from Windows to Microsoft 365 Copilot—the question remains how much actual revenue and profit the company is extracting from the arrangement.

Microsoft earns a cut from every query processed through ChatGPT that runs on Azure infrastructure. But building and maintaining the massive computing power needed to train and serve large AI models like GPT-4 is extremely expensive. Microsoft has not disclosed specific profit margins or return-on-investment figures from its OpenAI partnership, and the company’s recent comments suggest it is weighing whether the economics are sustainable.

Adding to the uncertainty, OpenAI’s GPT-4 has been criticized internally at Microsoft for being “too expensive and too slow” to meet the needs of mass-market consumer products.

Those concerns have reportedly led Microsoft to accelerate the development of its own in-house small language models—lighter, cheaper alternatives designed to support more nimble applications without relying entirely on OpenAI’s massive systems.

A Cooling Alliance?

The partnership has seen visible signs of cooling in recent months. Microsoft, which initially had exclusive cloud provider status for OpenAI, has since lost that exclusivity, though it retains a “right of first refusal” for future contracts. OpenAI has turned to Oracle and SoftBank as new collaborators in its expansion push, particularly as it embarks on building new data centers for AI development under its Stargate initiative.

Reports also surfaced that Microsoft canceled two major data center projects linked to OpenAI’s growing demand for computing, signaling Redmond’s unwillingness to further stretch its infrastructure to accommodate OpenAI’s increasingly aggressive roadmap.

In parallel, OpenAI CEO Sam Altman has grown more vocal about the company’s need to work beyond Microsoft. In a separate episode of The Circuit, Altman said although OpenAI has got a lot of great work with Microsoft, he thinks this is more than any one company can deliver, which means that OpenAI’s mission to reach Artificial General Intelligence (AGI) requires a broader base of support and collaboration.

Despite the increasingly public signs of divergence, Nadella insists Microsoft remains “thrilled” to have access to OpenAI’s models and continues to benefit from their success. “Having that multifaceted partnership is what we are really focused on,” he said. “Why would any one of us want to go upset that?”

He added: “Every day that ChatGPT succeeds is a fantastic day for Microsoft.” But the CEO’s remarks may also reflect a subtle recalibration: Microsoft is clearly preparing for a future in which OpenAI is no longer its sole, or even primary, AI partner.

Tech rivals, like Salesforce CEO Marc Benioff, have suggested Microsoft might eventually sever ties with OpenAI altogether, citing the growing cost and performance issues. Microsoft has so far avoided giving any indication of a formal split, but it is building out its own models, such as Phi-3, and doubling down on homegrown AI teams.

Nadella said he expects OpenAI to work with multiple partners going forward—a development he sees as natural.

However, the picture is changing fast. What began as a mutually dependent partnership has morphed into something far more fluid, as both Microsoft and OpenAI seek autonomy, scale, and profit in the fast-moving AI race.

This means that the next phase may not be about deepening ties with OpenAI, but about ensuring that the billion-dollar investment delivers real returns—or, failing that, building a competitive engine of its own.