We build AI systems that learn from real experience — agents that remember, adapt, and act on behalf of the people who use them.

Our research focuses on continual learning and collective intelligence. We develop our models through recursive self-improvement loops on Mint Recursive. Our infrastructure spans model-harness co-design, long-term memory, generative interfaces, and the post-training systems required to serve all of it at scale.

How AI systems autonomously improve their own capabilities through iterative learning and experimentation.

• Mint Recursive — autonomous research loops that turn ideas into runnable experiments and drive continuous model improvement.
• Mint Recursive Cookbook — collection of recipes for building self-improving AI systems with Mint Recursive infrastructure.

How agents learn from interaction over time, rather than only at training.

• LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget — architecture-aware execution stack for million-token RL post-training under a fixed GPU budget.
• Mint Recursive — RL infrastructure for experiential intelligence. Abstracts compute scheduling, distributed training, rollout, and evaluation behind a unified interface.
• MinT: Managed Infrastructure for Training and Serving Millions of LLMs — scaling LoRA-based post-training and serving across frontier-scale models and million-scale adapter catalogs.

How specialized models compose and route to handle complex, multi-faceted tasks.

• Macaron-V1.1 — agent model built on GLM-5.3, combining a shared base with specialized LoRA experts for Chat, Agent, Coding, and Generative UI through a Mixture-of-LoRA architecture.
• On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters — studies parameter-efficient fine-tuning as persistent local state, and how personal models scale up, down, and out on shared foundations.
• Mixture-of-LoRA Harness — serving harness for Macaron's Mixture-of-LoRA architecture, supporting specialized LoRA routing on shared base models.

How agents form, retrieve, and apply memory across sessions and over extended timelines.

• δ-mem — lightweight online memory mechanism that compresses historical information into compact state directly coupled with attention computation.
• MindClaw — experiments in personalized long-term memory through per-user LoRA fine-tuning.
• Streaming Memory Benchmark — evidence-based streaming evaluation for diagnosing memory formation, retrieval, application, latency, and token efficiency.

How agents create interfaces dynamically, rather than relying on fixed templates.

• Macaron-A2UI — generative UI model for personal agents. Generates natural language together with lightweight executable UI actions.
• UI4A — component-native generative UI harness. Lets agents generate code with imports, components, state, and actions inside enforced runtime boundaries.
• Macaron-A2UI-Bench — benchmark for controlled evaluation of generative UI capabilities in personal agents.

How models understand spatial relationships and make decisions from visual input.

• MJev-Spatial: System One Spatial Intelligence via Native Attention Matching — multimodal decision model built on Qwen3.5-4B. Uses the backbone's pretrained query and key projections to score candidates for choice, numerical prediction, and pointing tasks, without generating answer tokens.

How we measure agent behavior in long, complex, and culturally varied contexts.

• Macaron ChatBench v2 — evaluates conversational behavior across expanded real-world scenarios, using the production system prompt, user persona, and relevant conversation history.
• UI4ABench v2 — 60 UI-generation tasks inspired by real-world scenarios, evaluating first-attempt delivery, functional completeness, and visual quality.
• Macaron LivingBench — long-horizon benchmark for evaluating agents across weeks of interaction, daily-life scenarios, and different cultural contexts.
• Streaming Memory Benchmark — evidence-based streaming evaluation for memory formation, retrieval, application, latency, and token efficiency.

• Macaron-V1.1 — latest Macaron release, post-trained with Mint Recursive: a 752B release combining a 744B GLM-5.3 base with four 2B LoRA specialists. See the model card for benchmark results and evaluation protocols.
• Macaron-V1-Venti — previous-generation 748B-parameter flagship, post-trained from GLM-5.2 with four specialized LoRA experts.
• Macaron-V1-Tall — 50B-parameter Macaron-V1 model designed for local deployment.
• Macaron-V1-Coding-Venti — 744B merged checkpoint combining the GLM-5.2 base with Venti's Coding LoRA specialist for coding, terminal, and software-engineering workflows, without runtime LoRA routing.
• Macaron-A2UI — LoRA adapters for generating structured A2UI v0.8 cards in personal-agent interfaces. Model weights: Tall · Grande · Venti.
• MJev-Spatial-4B — multimodal model built on Qwen3.5-4B for spatial understanding, supporting choice, numerical prediction, and pointing.
• Macaron Model Previews — compare the latest Macaron models side by side and test them on real-world tasks.
• Mint Recursive Cookbook — turn new research ideas into runnable Mint Recursive recipes and agent-driven experiment loops.

• Website — research, releases, and updates.
• Publications — full list of papers and technical reports.
• Contact: [email protected]