MiniMax M2.7 just dropped and it is turning heads. The model scores top marks on coding benchmarks. It costs one tenth of what Claude Opus charges. And it can actually improve itself. This is not just another AI launch. This is a wake up call for the whole industry.
Here is a model that open source developers around the world are calling a game changer.
Just look at the price. You get Opus level quality for one tenth the cost. That is hard to ignore.
Some say the gap between open and closed models keeps shrinking. Every new release proves them right.
On PinchBench, a test for open source coding agents, M2.7 sits at number one.
It also tops the global open source coding agent leaderboard. Its stability score is way ahead of the older M2.5.
On OpenRouter, the annual cost of tokens for all global models has already passed one thousand billion dollars. And MiniMax M2.5 is one of the fastest growing models on the platform.
At the GTC conference, Nvidia announced a global open source project called OpenClaw. It is ten times faster than Linux.
Not to be outdone, Nvidia quickly followed up and launched NemoClaw based on OpenClaw.
Nvidia also praised OpenClaw as a dynamic claw. The reason is simple. MiniMax is the core contributor behind the OpenClaw ecosystem. It is one of the founding supporters of the project.
During GTC, MiniMax M2.7 made a splash. It demonstrated four core skills. Multi agent collaboration. Self evolution. Direct execution. And a stable agent system.
Beyond these demos, the first thing people noticed was the team itself. Engineers and researchers were showing off their own projects.
M2.7 has four major upgrades. Here is what makes it special.
Powerful Cowork Agent Model
The truth is, the M2.7 release made a lot of people happy.
The Agent Team and Coding model is not just a chatbot. It has native multi agent collaboration. It can use tools. And it can work on its own.
During live demos, the model showed it could read code. It could find bugs. It could fix them. It could write full programs. It could handle MLE tasks. It could build Android apps. And it could create games.
Office work is also a breeze. Excel, Word, and PowerPoint are all handled smoothly.
One line sums it up. M2.7 is a model that can actually get work done. It is a large model with real collaboration skills.
Multi Agent Collaboration Cowork Agent Model
MiniMax M2.7 is a model with native Agent Teams. It supports multi agent collaboration.
Under the hood, the model uses role identity, task data, adversarial reasoning, high level planning, and hidden state awareness. All of these become part of the model core.
Right now, the Cowork agent model organizes multiple skills into complex execution paths. It is very good at skill based execution.
To put it simply, M2.7 has a multi agent collaboration system. Each agent has a role. Each agent has data. And they work together on tasks. For example, one agent might write code. Another might test it. And a third might deploy it.
From a technical view, this is a complex workflow problem. The model plans execution paths. It assigns agent roles. It handles scheduling. And it manages the overall process.
More importantly, the model itself is the core. It is a Cowork Agent dynamic system.
Soon, M2.7 will launch a multi agent collaboration system with roles like researcher, predictor, designer, and more. Each agent has its own skills, tools, and memory.
Automatic tool use is also a strong point. Web applications are handled well.
There is another interesting test. In complex execution, does M2.7 perform well with external tools?
The answer is yes. The model provides UI skills, web search, external tools, and file systems.
After learning skills from GitHub, M2.7 released a new version. The visual effects and animations are stunning. And the core multi agent collaboration logic is stable.
The key idea is a loop. Experiment. Summarize skills. Improve. This is also how the team built MaxClaw. They tested and verified the path.
MiniMax and Nvidia launched MaxClaw together. It is a claw expert. In other words, an expert is a skill. And skills are the core of the system.
Right now, MiniMax has open sourced 6 key skills. The claw expert is already online. MiniMax invites users to create their own paths. The potential is huge.
ai nudifier
our dream ai
The claw expert is very popular. The agent investment team also joined in. They created a MiniMax M2.7 multi agent collaboration demo.
ai erotic smut
Here are some highlights from the agent investment team demo.
After launch, different agents with different roles look at the same task from their own angle. They share information within their scope.
Then they combine their views and reach a joint conclusion from different angles.
Finally, the team leader agent makes a decision. It creates a full investment report. The first agent view is complete.
Even stock analysis and investment research.
Strong Coding Skills That Surprise Everyone
In real world tests, M2.7 performs way above expectations.
In the past, M2 series models were known for open source work. But their coding skills were just average. They were not top tier.
In this new version, M2.7 has native code understanding and generation. It can read code. It can find bugs. It can fix them. It can build systems. It can write tests. And it can handle complex debugging.
When it comes to debugging, M2.7 is also very strong.
In a real system, when an API fails, the error message only gives partial info. The model needs to read logs. It needs to find key clues. And it needs to figure out the root cause. It also needs to find hidden issues.
With logs and code, M2.7 quickly finds the root cause. It says the missing SKU value for SKU-1003 is the problem. And it fixes it right away.
Plus, M2.7 can create a lobster game with a Shanghai flavor.
The user plays as a lobster. To succeed, the lobster must complete a series of actions on a web page.
The user controls the lobster through the web page. The lobster can move. It can jump. And it can interact with objects.
The game has three endings. Success means the lobster answers the sea and gets a happy ending. Failure means the lobster becomes bad ending. And there is a funny ending too.
Soon, M2.7 successfully created the game. And it added strong creative elements.
Office Automation Gets a Major Upgrade
In the new version, M2.7 also shines at Excel tasks.
For example, when given data from a Chinese company annual report, M2.7 can use Excel to create charts. It can do predictions. And it can do valuations.
The final output is a professional report. It has Excel charts. It has profit metrics. It has revenue metrics. It has debt structure metrics. It has cash flow metrics. And it has valuation metrics. Plus forecasts for 2025 to 2027.
For finance professionals, this is incredibly practical.
Click to view more
M2.7 Redefines AI Research Agents
AI research agents are becoming the new standard in the AI world.
In large model training, the gap between pre training and post training is closing. In other words, models can now optimize themselves.
But few teams have actually succeeded on this path.
Anthropic has already shown that model self evolution works. They used recursive self improvement. OpenAI has also pushed AI research agents. Google DeepMind has done the same through AlphaTensor, AlphaCode, Gemini 3 Deep Think, AlphaEvolve, and more. Recursive self improvement is now the new battleground.
The MiniMax team explored a new path with M2.7. They gave the industry a new reference.
This time, they did not use a large model to train a small model. Instead, they used a small model to train a large model. They built a system called Agent Harness.
The team used M2 series models across different sizes. From 1B to 4B. They wrote code. And they built a complete CI test and evaluation project called Agent harness.
The system includes a data pipeline. It has training code. It has evaluation tools. It has team collaboration. And it has persistent memory. The model can plug into the development process. It can read code. It can execute commands.
In practice, researchers only need to write prompts. They use an RL method to let the model handle most of the execution. This approach is simple. It is fast. And it works.
In the Chinese version, the agent reads the experiment state. It reads the logs. It finds the problem. It gives instructions. And it directly fixes the code. It submits the merge request. At the same time, it runs tests to verify.
Normally, this would need a whole team. But M2.7 handles 30 to 50 percent of the work on its own.
Researchers can plug M2.7 directly into the Agent Harness environment.
More precisely, M2.7 optimizes the internal scaffold code of a model.
In the full process, the steps are read failure trajectory, plan fix, modify scaffold, run test, compare results, and submit code. This loop runs 100 times.
In this process, M2.7 finds effective optimizations. It improves model efficiency. And it boosts performance by 30 percent.
M2.7 Proves Self Evolution Works
In open tests, the self evolution results also got solid proof.
The team used a simple framework. They gave the model time limits, search space, and optimization goals. Then they let M2.7 learn on MLE Bench Lite across 22 tasks.
At each step, the model reads the current state. It uses search to explore the space. It reviews historical info. And it creates an optimization plan.
In a 24 hour open test, M2.7 won 9 gold medals, 5 silver medals, and 1 bronze medal. The average medal rate was 66.6 percent. It beat Gemini 3.1. It was close to Opus 4.6 at 75.7 percent. And it was close to GPT 5.4 at 71.2 percent.
This proves that M2.7 has found a new way. In a closed system, it uses trial and error. It uses feedback. And it optimizes itself step by step.
This success proves that small models can learn from big models. And they can get better.
OpenRoom Creates Interactive AI Worlds
The most fun part is that M2.7 also supports interactive creative projects. MiniMax launched an open source project called OpenRoom.
It is an interactive web space. You say one sentence. And the scene changes in real time. The characters react. And the environment responds.
In this small world, each AI character has its own personality. They have their own paths. And they have their own memories.
It is like an AI research agent world. But it is also a virtual lobster world.
