Articles / Macaron-V1 Launches 2M-Context MoL Architecture

Macaron-V1 Launches 2M-Context MoL Architecture

23 7 月, 2026 • 4 min read • Macaron-V1MoL-Architecture

Macaron-V1 Launches 2M-Context MoL Architecture

Macaron-V1 Overview

A groundbreaking open-source AI model — Macaron-V1 — has been released by Xin Zhou Technology’s frontier lab Mind Lab, redefining personal intelligent agents through a novel Mixture of LoRA (MoL) architecture built atop GLM-5.2.

“Most ‘personal agent’ products still rely on plug-in chatbots — Macaron-V1 breaks that mold with purpose-built, isolated expert modules and native 2-million-token context support.”

🧠 Model Variants & Technical Foundation

Macaron-V1 ships in two production-ready variants:

✅ Macaron-V1-Venti (Flagship)

  • 748B total parameters: 744B base + four 1B LoRA experts
  • First-ever GLM-5.2 post-training model
  • Native 2M-context support — enabled via LongStraw training infrastructure
  • Optimized for cloud/API deployment

✅ Macaron-V1-Tall (Lightweight)

  • 35B parameters, fine-tuned from Qwen-3.6
  • Designed for local, privacy-first, cost-sensitive deployments

Both versions are fully open-weighted, accompanied by official APIs and tooling.

⚙️ MoL Architecture: Four Specialized Experts

Instead of monolithic parameter sharing, Macaron-V1 freezes the GLM-5.2 base and attaches four independent LoRA experts — each trained in isolation for domain-specific mastery:

Expert Role Key Capabilities
L0 Chat Dialogue & Instruction Following Natural language grounding, multi-turn coherence, safety-aware response generation
L1 Agent Personal Task Orchestration Long-horizon planning, tool chaining (e.g., calendar + email + finance), dynamic adaptation to user workflows
L2 Coding Code Generation & Execution Full-stack development (HTML/CSS/JS/WebGL/Three.js), rigorous PBR rendering, self-debugging via browser-integrated testing
L3 UI/A2UI Generative UI Rendering Real-time, interactive dashboard synthesis (e.g., financial control panels), live data-binding, animation-driven UI updates

💡 This modular design eliminates skill interference — e.g., coding logic never degrades conversational fluency — and enables plug-and-play capability extension.

🌟 Real-World Demonstrations

📊 UI4A Dashboard: “Personal Finance Control Center”

Using the Macaron Artifacts plugin (GitHub), users can generate rich, interactive dashboards via natural language:

Create a dark-mode personal investment dashboard with:
- Ring chart (asset allocation)
- Time-series line chart (net worth trend)
- Editable data grid (transaction log)
- Real-time联动: edit table → charts auto-update
- Macaron mascot sidebar with contextual financial advice

✅ Result: A production-grade, animated, glassmorphism-financed dashboard rendered live — with responsive data binding and zero external dependencies.

UI4A Dashboard

💻 L2 Coding: Steam-Punk Orrery & Ink-Scroll Landscape

Case 1: Mechanical Solar System (Three.js)

  • Fully self-contained HTML file
  • Procedural brass material (PBR), physics-based lighting, orbital mechanics
  • Mouse-drag orbit control + time-acceleration slider

Orrery Render

Case 2: Interactive Ink Scroll (WebGL)

  • 12-segment水墨 landscape with unique scroll physics per section
  • Procedural ink shaders, authentic guqin audio (Web Audio API) synced to scroll rhythm
  • Zero external assets — all geometry, texture, and sound generated in-browser

Ink Scroll

🤖 L1 Agent: FastAPI Issue Resolution

Macaron-V1 autonomously:
– Scanned tiangolo/fastapi GitHub repo
– Identified stalled PR #12481 (multi-model query/header/cookie support)
– Analyzed codebase architecture & style guide
– Implemented backward-compatible fix + 6 new tests + docs update
– Verified full test suite pass (3,000+ tests)

Result: Production-ready, merge-ready PR — demonstrating true engineering autonomy.

FastAPI PR

🧪 Benchmarks & Infrastructure

🔬 Proprietary Evaluation Suites

  • ChatBench: Measures honesty & self-awareness in long-agent contexts (e.g., “I don’t know” vs hallucination)
  • LivingBench: Simulated multi-week sandbox with dynamic noise, environment, and user behavior — tests recovery from deviation

📈 Cross-Benchmark Performance

Macaron-V1-Venti matches or exceeds top closed models (Opus 4.8, GPT-5.5) on:
– Personal Life Agent tasks (PinchBench, VitaBench)
– Generative UI synthesis (UI4A-Bench)
– Maintains competitive parity on coding (SWE-Verified, TerminalBench 2.1)

Benchmark Chart 1
Benchmark Chart 2

⚡ Enabling Infrastructure

  • LongStraw: Shared prefix caching cuts memory overhead — achieves 4.46M-token GRPO training on 32×H20 GPUs
  • MindForge: Unified RL training framework integrating Claude Code & Pi harnesses via HCP protocol
  • HyperRSI: Recursive self-improvement loop — generates harder tasks → audits trajectories → solidifies gains

🧭 Final Thoughts

Macaron-V1 doesn’t chase breadth — it pursues depth in personal agency. By decoupling capabilities into dedicated LoRA experts, enforcing strict context fidelity, and building evaluation around real-world longitudinal reliability, Mind Lab delivers a model architecture that feels less like an LLM upgrade — and more like the first true foundation for autonomous, trustworthy personal AI.

🔗 Official API: https://mintcn.macaron.xin/
📦 Open Source: Macaron-V1 GitHub (coming soon)
🧪 Try UI4A: Install Macaron Artifacts in Claude Code

Article originally published on “AI Cambrian” WeChat channel.