Macaron is a new kind of AI agent designed to enrich your everyday life. Simply tell Macaron what you need in plain English – plan a trip itinerary, start a fitness habit, journal your thoughts – and it will instantly build a mini-app tailored to your request. No coding or complex setup required. Interacting with Macaron feels like chatting with a friend who knows you, thanks to its unique persona and remarkable memory for context. As it learns your preferences over time, its assistance becomes increasingly personalized.

(PRUnderground) November 10th, 2025

What Makes Macaron Unique?

• Life-Focused Intelligence: Macaron is purpose-built for everyday life. Its hallmark capability is turning a single sentence into a working mini-app within seconds. This life-centric focus sets it apart from AI agents that mostly help with office work.
• Deep Memory & Personalization: Backed by a deep memory engine, Macaron builds a personal knowledge base about you. It remembers your likes, important events, and context from past conversations. That means it can remind you of tasks you mentioned, adapt to your style, and offer help that fits you specifically, without needing to be told twice.
• Collaborative Mini-App Creation: Macaron makes creativity social. Every mini-app it creates is easy to share or co-edit. You can even invite Macaron into a group chat to build something together in real time, making AI a collaborative experience. Macaron turns using an AI assistant into a shared activity, strengthening social bonds through shared innovation.
• Upcoming Feature – Daily Spark (Personal Inspiration): As a bonus, Macaron provides a Daily Spark – a short daily feed of inspiration and relevant content generated just for you. It might include news tidbits, motivational quotes, or fun facts tailored to your interests. It’s like getting a morning newsletter curated by an AI friend who understands your tastes.

Under the Hood: Technology Empowering Macaron

To deliver Macaron’s warm, personal experience, we built a cutting-edge AI foundation. Soon, our AI will become a trillion-parameter language model specialized for reasoning and conversation. Training and fine-tuning such an enormous model posed a serious engineering challenge. Our team innovated on multiple fronts – from how we distribute the model across hardware to how we fine-tune it efficiently – all to ensure Macaron’s “brain” is as smart and adaptable as it is friendly.

• Scaling a Trillion-Parameter Model with Hybrid Parallelism

Macaron’s AI foundation is a Mixture-of-Experts (MoE) model at unprecedented scale, combining hundreds of expert subnetworks. We devised a hybrid parallelism strategy to train this 1-trillion-parameter model without breaking the bank. Our training engine combines tensor, pipeline, expert, and sequence parallelism, dynamically adjusting each mode based on the scenario. For instance, it maximizes expert parallelism (using all experts across available GPUs) and only adds sequence parallelism for extra-long contexts, while keeping tensor parallel groups minimal to reduce communication. As a result, training the trillion-parameter model behind Macaron became feasible on our cluster, with wall-clock times and costs far lower than naive methods.

• Fast and Efficient Fine-Tuning with LoRA

Even with smart parallelism, fine-tuning a model of this size on specific behaviors would normally be extremely costly. Our solution is Low-Rank Adaptation (LoRA) – a technique that updates only small, low-rank matrices within the model instead of having to retrain all the parameters. By adding these tiny adaptation layers, we can train about 0.3% of the model’s weights to achieve the desired adjustments.

In our experiments, LoRA with a very low rank (between 8 and 128) achieved over 90% of the accuracy of full fine-tuning on our benchmarks, yet used only around 10% of the compute. Beyond that, returns diminished, so we settled on a LoRA setup that keeps Macaron’s model both powerful and efficient to update. This means we can frequently refine Macaron’s skills or personalize it further without retraining the entire giant model from scratch.

• Rapid Reinforcement Learning Alignment

Building Macaron wasn’t just about static training – it also required teaching the AI how to interact helpfully. We employed reinforcement learning (RL) to align the model’s behavior with human preferences (similar to how ChatGPT is fine-tuned with human feedback). Using a synchronized rollout-training pipeline, we dramatically accelerated this RL alignment process – slashing the time per training iteration by over 6× compared to naive methods. This allowed us to run many rounds of feedback-based tuning on Macaron’s model, honing its ability to ask clarifying questions, follow instructions faithfully, and respond with the right tone. As a result, we achieved the desired conversational improvements at roughly 10% of the usual training cost, meaning Macaron learned to be a helpful, friendly conversationalist faster and more efficiently.

Our journey in building Macaron’s AI engine yielded some general insights for AI research and engineering:

1. Hybrid parallelism for scale: Combining expert & sequence parallelism (with judicious tensor/pipeline use) enabled efficient training of the 1T-parameter model.
2. LoRA makes adaptation feasible: Low-rank fine-tuning let us adapt the giant model at ~10–25% of the usual cost, with minimal loss in performance.
3. Smarter tuning beats brute force: Adaptive techniques (per-layer or per-expert fine-tuning, improved RL loops) further boost model quality without growing its size.

Conclusion: From Advanced AI to an Everyday Companion

Turning cutting-edge AI research into an everyday companion required innovation at every layer. We combined user-centric product design with state-of-the-art model training to make an AI that is both highly advanced and deeply human-centric. The result is Macaron: an AI that leverages a trillion-parameter brain to deliver a warm, personalized touch in daily life.

Our technical breakthroughs in efficient training and model adaptation not only power Macaron today, but also pave the way for future improvements. For example, we are developing an automated hybrid scheduler (so the system can optimize its own parallel modes on the fly) and distilling Macaron’s knowledge into smaller models for personal devices. These efforts will ensure Macaron.im becomes even smarter and more accessible over time. We invite you to experience Macaron yourself – where a big leap in AI technology comes with a very human touch.

The post The Macaron Experience – AI That Helps You Live Better first appeared on

Original Press Release.