China’s 00‑Gen AI Team Unveils 748B LoRA Model Matching Opus 4.8 Performance

In July 2024, leading AI figures such as Mira Murati and Richard Sutton highlighted the need for models that can continuously learn after deployment. Mind Lab, a Chinese research lab founded by Liang Wenfeng and backed by a $50 million Series‑A round, responded by releasing Macaron‑V1, positioning it as a concrete realization of this “experience era”.

Model specifications : Macaron‑V1 is offered in two versions. The flagship Venti model contains 748 billion parameters—built from a 744 billion‑parameter GLM‑5.2 base plus four independent 1 billion‑parameter LoRA adapters—and supports a native 2 million‑token context window. The Tall variant is a 35 billion‑parameter model based on Qwen‑3.6, designed to run on a single Mac.

Mixture‑of‑LoRA (MoL) architecture treats the base model as shared knowledge and each LoRA as a distinct “personality” for chat, agent, coding, and generative‑UI tasks. At runtime a router dynamically selects the appropriate LoRA, making the specialization transparent to the user. This design reframes LoRA from a cheap fine‑tuning trick to a persistent, writable, and roll‑backable state that can evolve with usage.

Benchmark performance : In a comprehensive evaluation Macaron‑V1‑Venti ranked first on SWE‑bench (85.6 points), outpacing DeepSeek‑V4‑Pro (80.6), MiniMax‑M3 (80.5) and Kimi‑K2.6 (80.2). It also achieved higher scores than Opus 4.8, GPT‑5.5 and Gemini 3.1 Pro on most tasks, especially in everyday scenarios, code generation, and generative‑UI. On the Deep‑SWE long‑context benchmark it scored 58.4 versus the base GLM‑5.2’s 46.2, demonstrating a 12‑point gain from post‑training reinforcement learning.

Open weights and private deployment : Unlike many closed‑source competitors, Macaron‑V1’s weights are released under an open license, allowing customers to host the model on‑premises and keep all data private.

Continuous‑learning rationale : The article argues that pre‑training is a “data era” while the next frontier is an “experience era” where models improve through real‑world interaction. LoRA makes large‑scale reinforcement learning feasible: Mind Lab completed LoRA‑based RL on a 1‑trillion‑parameter MoE model using only one‑tenth the compute of full‑parameter fine‑tuning, achieving stable convergence.

Comparison with competitors : Thinking Machines Lab’s Inkling (975 billion parameters, $20 billion funding) still lags behind Chinese open‑source bases such as GLM‑5.2. Mind Lab’s strategy of standing on the shoulders of the Chinese open‑source ecosystem, rather than training a base from scratch, yields comparable or superior capabilities with far less capital.

Scaling insights : Experiments show a third scaling law—performance improves linearly with the number of cooperating models. Training 200 independent adapters on the same base raised accuracy from 36.44 % (single adapter) to 48.67 % (majority vote). Mind Lab’s MinT platform already indexes a million LoRA adapters, each sharing the same trillion‑parameter base, illustrating the feasibility of massive model‑of‑models collaboration.

Commercial traction : The API launched in early July generated $10 million ARR within two weeks, with pricing of $6 per million tokens for the flagship version. Early adopters include smartphone manufacturers, audio‑hardware startups, and major AI hardware firms.

Conclusion : The decisive factor in future large‑model competition will be the ability to keep models learning after deployment. Macaron‑V1 demonstrates that LoRA‑enabled reinforcement learning at the trillion‑parameter scale makes continuous improvement practical, turning a static language model into an ever‑growing “experience‑intelligent” system.