Mind Lab has released the Macaron-V1 model, featuring a Mixture-of-LoRA architecture with 748B parameters and four independent LoRAs, achieving a top score of 85.6 on the SWE-bench Verified benchmark, ranking first in code capability. This Chinese lab, founded less than a year ago, has reached technological parity with Silicon Valley teams backed by $2 billion, using only $50 million in funding. Its LoRA-based reinforcement learning approach has garnered support from industry leaders including Turing Award winner Richard Sutton and DeepSeek’s Liang Wenhong. Mind Lab also introduces the concepts of “continuous learning” and “collective intelligence,” asserting that the model can continue learning and collaborating in real-world environments after leaving the lab. Within just two weeks of commercialization, it achieved $10 million in annual recurring revenue (ARR), validating that this technical path addresses real-world pain points.

Article author and source: AI New Era

Last week, the AI community was discussing one thing, but most people only saw half of it.

On July 15, former OpenAI CTO Mira Murati launched Inkling, the first model from Thinking Machines Lab, in San Francisco.

$2 billion in funding, 975 billion parameters, a 100-person elite team—Silicon Valley’s most watched AI startup has finally revealed its cards.

Almost on the same day, Richard Sutton, the father of reinforcement learning and 2024 Turing Award winner, nearly seventy years old, announced the founding of Oak Lab.

The old man's exact words were: "Current deep learning methods are fragile and inefficient; what's needed is not patchwork fixes, but a complete restart from the ground up."

It's not just Sutton who thinks this way. Liang Wenheng, founder of DeepSeek, made nearly the same judgment.

He compared the development of AI to a series of steps: language models, CoT, agents... and after agents, the next essential step must be continuous learning.

His view is that the core capability of the next-generation model must be continuous learning; only then can it be called a next-generation model. If continuous learning is achieved first, general intelligence may become much easier to attain.

When a technical path is backed by a Turing Award winner, Silicon Valley’s most watched AI startup, and the founder of China’s most innovative large model, it is no longer merely a technical choice—it has become an industry consensus.

The tide has turned.

While everyone was watching Silicon Valley, on the other side of the Pacific, a Chinese lab founded less than a year ago quietly delivered a surprising result in the same field.

Mind Lab officially releases Macaron-V1.

Mind Lab was founded in October 2025 as a cutting-edge research laboratory under Mindverse. The team consists of over thirty members, hailing from xAI, DeepMind, DeepSeek, ByteDance Seed, MIT, and Tsinghua University. In early 2025, the lab completed a $50 million Series A round led by Meituan, with participation from Ant Group, Sequoia Capital China, ZhenFund, and others.

Please note this amount: $50 million—it will be referenced again later.

748B parameters directly using reinforcement learning

Four LoRAs, each handling its own task

First, look at the model.

Macaron-V1 comes in two versions:

• Venti Pro, 748B parameters, built on a 744B GLM-5.2 base with four 1B LoRAs, natively supports 2M context length;
• Standard version Tall, 35B, post-trained on Qwen-3.7, capable of running on Mac.

But the key isn't in the parameters—it's in those four LoRAs.

In the past, LoRA was the budget-friendly alternative to full-parameter fine-tuning—cost-saving, but with compromises.

Mind Lab has redefined it: the base model is the "commonality," and LoRA is the "individuality."

What does it mean? Just like all iPhones use the same iOS system, the apps you install, your typing habits, and your most frequent actions make your iPhone uniquely "your iPhone."

Conversational chat needs to be natural and empathetic, programming requires strict precision, agents require multi-step planning, and generative UIs need thoughtful interaction design—training them all within a single set of parameters causes them to conflict.

Macaron-V1's approach is simple and direct: break it down.

Four independent LoRAs, each dedicated to chat, Agent, programming, and GenUI, trained, evaluated, and rolled back separately.

Dynamic router switching at runtime—say “Help me schedule my day,” and it uses the Agent LoRA; say “There’s a bug in this code,” and it switches to the Coding LoRA. Completely transparent to the user.

This architecture has a formal name: Mixture-of-LoRA (MoL)—different capabilities are trained independently on a shared base, combined flexibly, and invoked on demand. Instead of forcing a single model to become an all-rounder, it enables a team of specialists to work together.

In terms of performance, Macaron-V1 has not yet reached the level of the strongest closed-source models across all capabilities. However, across the full suite of evaluations, it already matches or exceeds Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on most tasks—particularly excelling in everyday scenarios, coding ability, and generative UI.

SWE-bench Verified—One of the most rigorous code proficiency evaluations in the industry—Macaron-V1-Venti ranked first with a score of 85.6, outperforming the second-place model by 3 points and surpassing DeepSeek-V4-Pro (80.6), MiniMax-M3 (80.5), and Kimi-K2.6 (80.2).

Even more impressive is Deep-SWE, specifically designed to evaluate long-range, complex software engineering tasks. Macaron-V1-Venti scored 58.4, while the second-place finisher was its base model, GLM-5.2, at 46.2. Using the same base model, fine-tuning with LoRA reinforcement learning resulted in a direct 12-point increase. This is the incremental value that post-training can deliver.

More importantly: open weights with private deployment options. No matter how powerful a closed-source model is, your data still passes through someone else’s server. With Macaron-V1, you can bring the model back home—not a single byte of your data needs to leave your control.

During his WAIC talk, Andrew demonstrated a case: a user takes a photo of a scrambled Rubik’s Cube and sends it over; the model identifies the state, solves it algorithmically, and generates an interactive 3D restoration guide—requiring the coordinated use of visual understanding, algorithms, agent tool invocation, and GenUI.

Another easily overlooked but highly robust design: Model and Harness are designed together.

What does this mean? Most model training and actual deployment occur in two separate environments—the tools and workflows used during training don’t match the real-world environment after deployment, leading to reduced performance.

From the start, Macaron-V1 trained and optimized the model alongside the production environment's harness—how tools are invoked, how the terminal is operated, and how long-running tasks are executed during training is identical to how they are handled after deployment.

The accompanying REPL harness enables the model to combine tools via code, save intermediate states, and reuse effective workflows—eliminating the need to start from scratch each time and reducing redundant calls and context waste.

This is not just a chat model—it's an intelligent agent that can get things done.

Why is it one of only two globally?

You might say: What’s so special about LoRA fine-tuning? Everyone’s doing it.

The remarkable aspect lies in two words: scale.

Many teams can perform LoRA reinforcement learning on a 30B model. But when the parameter scale reaches the trillion level, the difficulty increases exponentially.

The core challenge is training-inference inconsistency. In reinforcement learning, training and inference must occur simultaneously to collect feedback. When a trillion-parameter MoE model is sliced and distributed across dozens of GPUs, even minor discrepancies between training and inference precision cause gradient drift, leading to immediate non-convergence of the training curve.

For example: In a 30-person chamber orchestra, a slight pitch deviation can still be managed. But in a 100-person symphony orchestra, if even one instrument is out of tune, the entire performance becomes a disaster.

Only two teams worldwide can run LoRA reinforcement learning at the trillion-parameter scale.

One is Mira Murati’s Thinking Machines Lab. The other is Mind Lab.

There is no third party.

In December 2025, Mind Lab successfully implemented LoRA reinforcement learning on Kimi K2, a massive sparse MoE model with a total parameter count of 1.04 trillion. The training achieved stable convergence with only one-tenth the compute required for full-parameter fine-tuning.

This month’s LongStraw pushed reinforcement learning fine-tuning to 2M context length under fixed GPU compute. Macaron-V1 is the culmination of these advancements.

Mind Lab isn’t the only one on this path. John Schulman, Chief Scientist at TML and co-founder of OpenAI and inventor of the PPO algorithm, published "LoRA without Regret" in 2025, demonstrating through experiments that LoRA performs identically to full fine-tuning in most post-training scenarios. He referred to this as the “low-regret zone.”

The industry is also voting with real money. Fireworks AI, which recently completed a $1.5 billion Series D round, is valued at $17.5 billion and can host hundreds of LoRAs; Together AI raised $800 million in its Series C round and has made LoRA the default fine-tuning method.

Silicon Valley spent $2 billion on its foundation; it spent one-tenth of that to stand on the shoulders of giants.

Here is a very interesting comparison.

Mind Lab and TML follow the same technical approach, but their underlying strategies are completely different.

Due to various reasons, TML chose to train its foundational model from scratch. Inkling, with 975 billion parameters and 45 trillion tokens, is still, as Murati herself publicly acknowledged: “Not the strongest model available today.”

Evaluations show that Inkling is significantly weaker than Chinese open-source models such as GLM and Kimi. Building the foundation from scratch consumed a tremendous amount of their time and computational resources.

Mind Lab does the opposite—it doesn’t train its own base model but instead stands on the shoulders of China’s open-source model ecosystem.

Kimi K2 validates LoRA RL with trillions of parameters; V1-Preview is based on GLM-5.1, the official Venti version uses GLM-5.2, and Tall uses Qwen-3.7—each base model upgrade raises the starting line for post-training.

Chinese open-source models have reached near-SOTA levels globally. By taking these near-top-tier foundations and further enhancing key capabilities through LoRA fine-tuning, we can push them to true SOTA.

You don’t need to grow your own wheat to bake the best bread.

How harsh is the comparison?

TML: $2 billion in funding, 100-person team, NVIDIA gigawatt-scale computing partnership—Inkling’s capabilities still lag behind GLM-5.2.

Mind Lab: Over thirty people, less than $50 million in funding—developed stronger models across multiple dimensions including Chat, Agent, and Coding.

On the matter of LoRA reinforcement learning, Mind Lab is likely the team with the highest return on investment globally—without exception.

Continuous Learning: Why Do Models Become Smarter with Use?

After discussing the technology and comparisons, we arrive at the most critical question: What is Mind Lab really playing for?

The answer is hidden in the title of an article co-authored by Richard Sutton and David Silver, former chief researcher of AlphaGo: Welcome to the Era of Experience.

Pre-training is the 'data era': models learn from existing human-generated text.

The next era is the "Experience Era": AI's capabilities will no longer come from existing data, but from agents' long-term actions, feedback, and continuous learning in real-world environments.

The fundamental contradiction with large models today is this—they get frozen after training.

A user corrected the same error ten times, and next time, the model still has to reread the chat history. A Coding Agent repeatedly fails on the same codebase, but failure doesn’t become a new capability. The model processes millions of tasks daily, yet all experience is lost.

It is merely an infinite reasoning engine that repeats itself, not an evolving intelligent agent.

What Mind Lab aims to do is break this cycle. Their goal is to build an "experience intelligence machine"—a model that continues learning in the real world even after leaving the lab.

Experience becomes new capabilities, and those capabilities enable solving harder problems, which in turn generate richer experience.

Intelligence is no longer the result of a single training session, but becomes a continuous growth process.

Here, LoRA is no longer a "cost-saving tool," but rather a persistent state that can be written to, rolled back, and evolved independently.

The base carries general knowledge; LoRA carries your preferences, your coding style, and your business logic. It’s not a static patch—it’s memory that grows with interaction.

Experiments show that a lightweight adaptation module compressed to less than 0.5% of the base model's parameters remains stable and reliable—the stronger the base model, the more effective the small-scale updates become.

Mind Lab has been on this path for three years. In 2023, founder Andrew collaborated with Karthik Narasimhan, author of GPT-1, and Yao Shunyu to publish FireAct—the first work to enhance agent capabilities through parameter learning. After encoding agent trajectories into parameters, execution time dropped from 9.0 seconds to 2.7 seconds per run.

Pay a one-time cost for learning, and every subsequent use yields a return. From FireAct to Macaron-V1, this logic has been scaled up to trillions of parameters.

Swarm Intelligence: The Third Scaling Curve

Continuous learning is only half the story. The second core idea embodied by Macaron-V1, and perhaps the more disruptive one, is collective intelligence.

Each model continues learning based on its own experiences, causing different instances to evolve in distinct directions. Different user feedback, tasks, and data shape unique capabilities. Starting from the same foundation, over time, they grow into "different individuals."

Is this diversity a bug or a feature? The results were explosive.

Starting from the same Qwen3-30B base, the training objectives and algorithms are identical, with only the data order and masking changed.

Single adapter accuracy on AIME24: 36.44%. Majority vote with 198 different adapters: 48.67%. Repeated sampling 198 times with the same adapter yields a maximum of only: 43.78%.

Different "experiences" bring complementary strengths that a single model cannot provide. This isn't just "trying multiple times"—it's a genuine collaborative effect arising from diverse learning pathways.

Andrew calls this the third scaling curve at WAIC.

Rule 1: GPT-3-style Scale Parameters—Larger parameters lead to higher intelligence.

Article 2: DeepSeek R1-style Scale Thinking Tokens—More reasoning tokens enable higher intelligence.

Article 3: Scale Number of Models — Greater collaboration among more models leads to a higher ceiling of intelligence.

In the experiment, the number of models increased from a few to 200, and performance improved linearly on a logarithmic scale. If this trend holds, the next step is to scale from 200 to 20,000, then to 2 million.

Mind Lab's proprietary MinT platform has established a directory of over a million addressable LoRAs, with one million LoRAs sharing a single trillion-parameter base model.

Note—this is not one million small models; it is one million large models with one trillion parameters each.

Humans share over 99% of their genome, but nearly identical biological foundations do not produce identical minds. It is the diversity of intelligence and its collaboration that have led to humanity’s greatest achievements.

$10 million ARR in 2 weeks: Fastest commercial validation

No matter how impressive the technical roadmap, if the market doesn’t adopt it, it’s just a PowerPoint. Mind Lab opened its API for commercial use in early July.

Result: $10 million in annual recurring revenue after two weeks.

This is likely the fastest record of any model company achieving $10 million in ARR since the first model's release.

Numbers matter, but speed matters more. The fact that customers have voted with real money so quickly on this technical path shows that it solves real, pressing problems.

Mind Lab doesn't just sell tokens. Its core business logic is that customers are purchasing the ability for models to continuously learn and iterate within their own use cases.

Phone manufacturers train product knowledge into dedicated LoRAs, while headphone companies embed interaction preferences into model states—core data doesn’t need to be sent to base model providers for retraining; customers retain full control.

ShanYin Technology, AI hardware startup Odyss, and a leading smartphone manufacturer are already on the collaboration list.

The API call structure also reveals the issue: 20% chat, 30% agent workflows, 30% code generation, and the remainder from GenUI—scenario distribution is balanced, not fueled by a single viral hit creating false growth.

Premium plan: $6 per million tokens, Standard plan: $4, Ultra plan: $2.

Mind Lab's vision is simply: If you can use a token, you can train a token.

Training a model should ultimately be as simple as invoking a model—no need for extensive pre-sales support or customized manpower. Continuous learning itself is a scalable model capability.

The door opened.

Richard Sutton declared the "Age of Experience" in Toronto. Liang Wenheng believes that only models with continuous learning capabilities can be called the next generation. Mira Murati built a trillion-parameter LoRA reinforcement learning infrastructure in San Francisco. Fireworks AI is valued at $17.5 billion.

Everyone is betting on the same direction: the model should continue learning after deployment. LoRA is the key technology that makes this feasible at the trillion-parameter scale.

Mind Lab is the youngest player in this space, but its starting posture may be the most efficient—without a foundational base, it leverages China’s open-source ecosystem, achieving technical capabilities on par with Silicon Valley teams backed by billions of dollars, with just over thirty people and less than $50 million in funding.

Behind this is a simple but profound judgment:

The ultimate competition in large models isn't about who trains the largest foundation model, but who can make their models continuously smarter.

Pre-training is a one-time infrastructure investment. Continuous learning is the true moat.

Mind Lab aims to build an experience-driven intelligent machine—a model that continues to learn in the real world even after leaving the lab, becoming smarter the more it’s used. When enough of these intelligences begin to collaborate, a new path toward higher intelligence will open up.

Murati pushed this door open in San Francisco with $2 billion and a team of 100.

First to crack open the door a little was a young laboratory on this side of the Pacific, standing on the shoulders of China’s open-source ecosystem.