Beyond Scaling: How Macaron‑V1 Opens a New Path for Continuous Learning in Open‑Source AI

In June 2026, Zhipu released the large‑scale MoE model GLM‑5.2, which achieved strong performance but, like most foundation models, becomes static once training ends and parameters are frozen. Zhipu’s roadmap highlights memory, continual learning, and self‑evaluation as the next challenges, suggesting that scaling alone is no longer the sole direction.

Macaron‑V1: A Model That Grows After Deployment

Mind Lab’s open‑source Macaron‑V1 builds on GLM‑5.2 (and a smaller Qwen‑3.6‑based variant) with native 2 M token context. Its core innovation is to enable continuous learning and large‑scale collaboration among many LoRA adapters, directly addressing the post‑training growth problem.

The model emphasizes two keywords: adaptation (continuous learning) and collaboration (group intelligence). Mind Lab argues that the decisive factor for the next generation of AI is not merely enlarging a single monolithic model but combining recursive self‑improvement (RSI) —post‑training loops that generate experience and refine behavior—with multi‑agent collaboration , where specialized agents share a common infrastructure.

Benchmark Results

Macaron‑V1 consistently outperforms its GLM‑5.2 base on evaluated tasks. Notably, on the UI4A benchmark that tests front‑end design ability, Macaron‑V1 scores 87.7, far surpassing Claude Opus’s 4.8. In a practical test, a single simulated macOS web page generated with Claude Code and Macaron‑V1 took 22 minutes, producing a 67 KB HTML file with fully functional calculator, photo viewer, weather, browser, and translation features.

These gains stem from a handful of LoRA adapters. The model’s Mixture‑of‑LoRA (MoL) architecture groups tasks with similar skills into the same LoRA, allowing them to reinforce each other, while divergent tasks are assigned to separate LoRA experts to avoid interference.

MoL Architecture Details

Macaron‑V1 mounts four 1 B LoRA experts on a frozen GLM‑5.2 base:

L0 Chat : core dialogue and instruction following.

L1 Agent : heavy tool use and long‑range complex tasks.

L2 Coding : code understanding and terminal interaction.

L3 GenUI : generative UI rendering and UI‑driven operations.

The flagship “Venti” variant totals 748 B parameters (744 B base + four 1 B LoRAs). A smaller “Tall” version contains 50 B parameters (35 B base + four 3.7 B LoRAs based on Qwen‑3.6).

LoRA as Persistent Local State

Traditionally, LoRA and other PEFT methods are viewed as cheap substitutes for full‑parameter fine‑tuning. Mind Lab repositions adapters as a persistent local state : the shared base provides generic capabilities (the “commonality”), while adapters encode individualized behaviors such as user preferences, tool‑use habits, or memory updates (the “personalities”). This perspective is supported by their “On the Scaling of PEFT” study, where LoRA‑based reinforcement learning on a trillion‑parameter MoE model reduced compute and communication costs to roughly 10 % of full‑parameter RL while maintaining stable training curves.

Further experiments show that, under comparable RL budgets, a 1.5 B model with full‑parameter RL gains 8.33 % normalized improvement, whereas a 7 B model using LoRA (r=64) achieves 11.31 % improvement, and a 32 B model with LoRA (r=8) reaches 20.61 %—demonstrating that smaller trainable fractions can yield larger gains on stronger bases.

Group Intelligence via Adapter Collaboration

Macaron‑V1 explores whether multiple LoRAs can cooperate. Experiments on a Qwen‑3‑30B base show that a single adapter attains 36.44 % accuracy on AIME24, while a majority‑vote ensemble of 198 adapters reaches 48.67 %. Re‑sampling the same adapter repeatedly only improves to 43.78 %, indicating that diverse learning trajectories provide complementary information beyond simple redundancy.

Beyond voting, Macaron‑V1’s MoL enables dynamic routing: different LoRA specialists can be selected per task phase, allowing a complex workflow (planning, coding, retrieval, UI generation) to be handled by the most suitable expert.

Infrastructure: MinT, MindForge, and LongStraw

To support million‑adapter scale and reinforcement learning on trillion‑parameter models, Mind Lab built the MinT platform (Managed Infrastructure for Training and Serving Millions of LLMs, arXiv:2605.13779). Key capabilities include:

Adapter‑only communication between Actor and Learner, avoiding full‑model weight transfer.

Support for a million‑scale adapter catalog, each with independent identity and version history.

End‑to‑end handling of models up to a trillion parameters.

On top of MinT sits MindForge , a training framework that integrates the same Harness Context Protocol (HCP) used in production, ensuring alignment between training roll‑outs and serving conditions. MindForge implements a recursive self‑improvement (RSI) loop: generate harder seed tasks, evaluate learning value, roll out trajectories, audit, and iterate, finally solidifying the experience into model parameters.

The LongStraw system addresses the mismatch between long‑context inference (millions of tokens) and reinforcement learning, which traditionally caps at 256 K tokens. By treating shared prompts as non‑gradient state and serially replaying response branches, LongStraw enables deterministic execution of 2,097,152 token positions on a 32‑GPU H20 cluster, making million‑token RL feasible under a fixed GPU budget (arXiv:2607.14952).

Model‑Integrated Harnesses

Macaron‑V1 also integrates three harness components during training:

UI4A : a generative UI layer that retains native HTML flexibility and can import libraries from NPM or arbitrary URLs, allowing the model to produce rich interactive content without task‑specific fine‑tuning.

REPL Harness : provides a persistent Python namespace where the model can write code, chain dependent steps, and save verified tools for reuse across sessions.

HCP (Harness Context Protocol) : a unified schema that packages all agent‑harness configurations (agents, skills, hooks, system prompts) for consistent use in both training and production.

Conclusion

Mind Lab has released the open‑weight variants Macaron‑V1‑Venti , Macaron‑V1‑Coding‑Venti , and Macaron‑V1‑Tall on Hugging Face, along with an official API for rapid experimentation. While the vision of a fully realized Experiential Intelligence Machine remains a long‑term goal—challenges such as delayed feedback, attribution difficulty, self‑reflection errors, and stability of continual updates persist—Macaron‑V1 demonstrates a concrete step: a strong open‑source base, plug‑in LoRA‑based continual learning, collaborative multi‑adapter intelligence, and infrastructure capable of handling millions of adapters.

Future work will need to address feedback latency, attribution, efficient self‑learning, and stable continual updates, but the current progress suggests that experience‑driven AI is becoming increasingly attainable.