Macaron-V1: Mixture-of-LoRA for Agentic AI

On 21 July 2026, Mind Lab released Macaron-V1, the first official release of their Mixture-of-LoRA (MoL) architecture for agentic AI. The flagship variant, Venti, combines a 744B-parameter GLM-5.2 base model with four 1B-parameter LoRA specialists — one each for chat, agent, coding, and generative UI tasks.

The Mixture-of-LoRA Architecture

Macaron's core architectural insight is that different AI capabilities benefit from specialised parameter allocations, but full fine-tuning separate models for each capability is prohibitively expensive. MoL addresses this by attaching small, swappable LoRA adapters to a frozen base model, with a router that selects the appropriate specialist for each incoming request.

Component	Venti	Tall
Base model	GLM-5.2 (744B)	Qwen 3.6 (35B)
LoRA specialists	4 × 1B	4 × 3.7B
Total parameters	~748B	~50B
Post-training pipeline	MinT + MindForge	MinT + MindForge
Context (training)	2M tokens (LongStraw)	2M tokens (LongStraw)

The four specialists are:

• L0 Chat: conversational backbone and instruction following
• L1 Agent: heavy tool use across long-horizon, dynamic, and complex tasks, including personal-life scenarios
• L2 Coding: code understanding, SWE tasks, terminal use
• L3 GenUI: UI4A rendering and UI-driven action

LongStraw: Multi-Million Token RL Training

Macaron-V1 introduces LongStraw, a training infrastructure that makes multi-million-token reinforcement learning practical. The technique evaluates the shared prompt once into a reusable resident state, then replays only the response branches that actually learn, accumulating gradients and performing one distributed update. This turns prompt length from a full-sequence memory problem into a state-lifetime problem.

Configuration	Max Context (GRPO)	Max Context (Resident Prefix)
8× H20 + Qwen3.6-27B	2.1M tokens	4.46M tokens
32× H20 + GLM-5.2 (full)	2.1M tokens	—

This makes multi-million-token RL training practical under a fixed GPU budget — a critical enabler for agentic models that must reason over long conversation histories and large codebases.

Benchmark Performance

Macaron-V1-Venti matches or outperforms frontier closed models across personal intelligence benchmarks, while remaining open-weight and self-hostable.

Benchmark	Macaron-V1-Venti	Claude Opus 4.8	GPT-5.5
VitaBench (personal agent)	Leader	—	—
PinchBench	Competitive	—	—
SWE-Verified	Strong	Strong	Strong
TerminalBench 2.1	Competitive	Competitive	Competitive

Coding is the area where Venti sits close to rather than ahead of the frontier, which the team identifies as a primary direction for future iterations. Personal intelligence — the ability to manage long-running, context-rich, multi-step personal and professional tasks — is where Macaron's MoL architecture provides the clearest differentiation.

Enterprise Deployment

The Venti variant requires approximately 748B parameters in memory. The LoRA specialists add only 4B parameters to the base model, meaning the incremental cost of multi-capability support is negligible relative to deploying the base model alone. The Tall variant at 50B parameters is designed for local deployment on consumer-grade hardware.

Macaron-V1 is available as open weights on Hugging Face, with a hosted API at mint.macaron.im. The coding-specialised build, Macaron-V1-Coding-Venti, ships with the coding LoRA merged into the base for direct use without routing.

Strategic Position

Macaron-V1 represents a different philosophy from monolithic models: rather than training one model that does everything moderately well, it trains specialised adapters that excel in their domains, routed by a lightweight inference-time selector. For enterprise teams, this means:

• Task-specific fine-tuning without full retraining. New capabilities can be added by training additional LoRA specialists against the same frozen base.
• Capability isolation. A poorly performing specialist can be replaced without affecting other capabilities.
• Resource efficiency. Four specialists share the same base model, requiring approximately the same inference memory as the base alone.