MindLab has released Macaron-V1, a model fine-tuned using the Mixture-of-LoRA method. This development is built upon the foundation of Zhipu AI GLM 5.2 and demonstrates how targeted reinforcement learning (RL) fine-tuning can significantly boost the capabilities of an existing model with minimal computational overhead.
The flagship Venti version, boasting 748 billion parameters and supporting a million-token native context, utilizes four independent LoRA expert adapters with 1 billion parameters each. These experts cover chat, agents, coding, and user interface domains. Each expert was trained in an isolated adapter space using the proprietary Mixture-of-LoRA technique, which prevents the interference between capabilities often observed during multi-task fine-tuning.
The MoL method represents a significant departure from traditional fine-tuning. Standard multi-task fine-tuning suffers from the oscillation effect, where improving one function leads to the degradation of another. MindLab's solution places numerous independent LoRA experts atop a frozen base model, with each optimized in its own isolated adapter space. A router mechanism selects the appropriate expert for each query, and new capabilities can be added simply by connecting another LoRA expert without retraining the existing ones.
The technical innovations in Macaron-V1 extend beyond MoL. The model incorporates LongStraw technology, which extends GLM 5.2's native context to two million tokens by reusing common prompt segments as a persistent state, substantially reducing the overhead associated with long sequences. Furthermore, there is a lightweight Tall version with 35 billion parameters designed for local deployment based on Qwen 3.6.
To support the system, three components were jointly trained: GenUI, responsible for displaying interactive visualizations beyond text output; REPL Harness, providing a persistent Python execution environment that allows models to create and advance tools for global utilities across sessions; and the MinT training infrastructure, specifically developed for MoL, where the Actor and Training Agent exchange only adapters, not full weights, supporting millions of adapter catalogs at a trillion-parameter scale.
MindLab also created two proprietary evaluation datasets. Macaron ChatBench assesses agent honesty in long-context scenarios, checking whether the model admits knowledge gaps instead of fabricating answers. Macaron LivingBench simulates a dynamic sandbox with three interacting mechanisms: dynamic noise, a dynamic living environment, and a dynamic user, measuring the model's ability to continuously understand user needs, verify information accuracy, replan task paths, and complete tasks before user patience expires.
On standard tests, Macaron-V1 achieves a generation speed of 320 tokens per second, significantly faster than comparable models, although this may reflect low inference load during early testing. Macaron-V1 proves that Chinese artificial intelligence is defined not only by the scale of the base model training but also by the efficiency of subsequent fine-tuning. MindLab's achievement using 64 GPUs confirms the feasibility of specialization while maintaining computational efficiency. As the quality of base models converges, the competitive advantage shifts towards the post-training layer, where MoL methods may define the ecosystems attracting the most capable applications.
A joint research group from Peking University and the Shanghai Institute of Microsystems, Chinese Academy of Sciences, has developed the world's first millisecond-level neuromorphic dynamic chip based on phase-change memristors. This development was published in the journal Science on July 14th.
This chip, measuring only 0.28 square millimeters, is capable of achieving a computation latency of 2.12 milliseconds. It integrates in-memory computing arrays for matrix operations and stepped shift arrays for adaptive integration.
The chip fundamentally changes the approach to computation in neural dynamics systems. The traditional Von Neumann architecture separates memory and processor, forcing data to constantly move between them, creating bottlenecks in latency and power consumption when solving differential equations required for neural dynamics. The Peking University team solved this problem differently by using the physical properties of the phase-change memory devices themselves to perform computations, instead of creating faster digital circuits.
Phase-change memristors possess continuously variable conductance, which can be precisely programmed and predictably evolve over time. Researchers mapped the search for adaptive integration required by neural dynamics solvers directly to the evolution of the conductance of these memristors. This means that the search, evaluation, and adjustment of step size—which digital circuits perform through many clock cycles—occurs via physical evolution within the device itself.
Furthermore, multi-level control of conductance allows for the creation of high-density in-memory computing arrays that simultaneously store neural network weights and perform matrix operations within a single array.
Experimental data demonstrates impressive performance. Compared to the most advanced ASIC accelerators, the system provides an acceleration ranging from 3.82 to 36.27 times while reducing power consumption by 3.9–7.8%. In tasks involving high-precision reconstruction of the brain cortex surface, the system outperforms the NVIDIA A100 GPU by 50.38–478.18 times. This marks the first instance where neural dynamics hardware has been brought to millisecond-level real-time operation, enabling applications previously limited to offline processing to become online real-time capabilities.
The article published in Science also contains a review paper describing this work as a paradigm shift in physically controlled computing. The significance of this development extends beyond demonstrated brain modeling and includes potential for real-time brain-computer interfaces, digital brain twins, neural navigation for surgery, and intelligent diagnosis of neurodegenerative diseases. The chip, operating at a frequency of 50 MHz with 9-stage pipelining, shows that architectural innovations can provide orders of magnitude improvement compared to a 'brute force' approach using GPUs for specialized computational tasks.
The research was supported by the New Cornerstone Investigator program, the National R&D Program, and the National Natural Science Foundation, as well as the Guangdong Key Laboratory of In-Memory Computing Chips. This chip represents a distinctly Chinese contribution to the developing field of physically controlled in-memory computing, where performance enhancement is driven by device physics rather than Moore's Law scaling.
