MiniMax Throws a Very Large Hat Into the Ring
China’s AI race just got louder, heavier, and a little more dramatic.
MiniMax, one of China’s fast-rising artificial intelligence startups, is reportedly working on a massive new large language model with 2.7 trillion parameters. That is not a typo. Trillion. With a “t.” According to reports citing people familiar with the project, the model is internally known as M3 Pro, though that name may change before launch. MiniMax also plans to release it as an open-weight model, potentially as early as the third quarter of 2026.
That makes this more than another “our chatbot got smarter” announcement. If MiniMax follows through, it could release one of the largest open-weight AI systems ever made available, and perhaps the largest from a Chinese company so far. Reuters, via The Economic Times, reported that the model may even become the largest open-weight AI model in the world, though the company declined to comment publicly.
The punchline is simple: MiniMax wants to go big. Not “slightly bigger.” Not “marketing-slide bigger.” Big-big. Godzilla wearing a GPU necklace big.
And because it plans to open-source the model, the impact could stretch far beyond China.
What Does 2.7 Trillion Parameters Actually Mean?
Parameters are the internal values an AI model learns during training. They help the system recognize patterns, predict text, follow instructions, reason through problems, and generate answers.
Think of parameters as adjustable knobs inside the model. During training, the system keeps turning those knobs until it becomes better at predicting what comes next. More parameters do not automatically mean better intelligence. That point matters. A bloated model can still perform badly if the training data, architecture, or tuning choices are weak.
Still, parameter count gives us one clue about scale. And 2.7 trillion is enormous.
MiniMax’s current flagship model, M3, reportedly has 428 billion parameters. The planned M3 Pro would be roughly six times larger. That is not a casual upgrade. That is the AI equivalent of going from a city bus to a cargo ship and saying, “Yes, this should fit the passengers.”
Reports say larger models tend to perform better on complex reasoning and multi-step instruction tasks. That does not guarantee MiniMax will beat every rival. But it explains why the company appears willing to chase scale. The frontier AI race has moved from simple chat to planning, tool use, coding, research, agents, and long-horizon problem solving.
That is where size can matter.
Open-Weight Is the Real Firecracker
The biggest detail is not just the size. It is the release strategy.
MiniMax reportedly plans to make the model open-weight. Open-weight models allow developers to download, run, modify, and customize the underlying model more freely than closed systems. That makes them very different from proprietary AI products that users can only access through an app or API.
This matters because open-weight models spread fast.
A capable open model can land in startups, universities, enterprise servers, coding tools, customer support systems, local devices, and experimental apps. Developers can fine-tune it. Companies can host it privately. Researchers can inspect it more closely. Smaller teams can build products without paying premium API costs to a closed-model provider.
That changes the economics.
Open-weight AI does not just compete on performance. It competes on control. A business may accept slightly weaker performance if the model costs less, runs privately, and avoids vendor lock-in. That is the quiet knife in this story.
US frontier labs often rely on expensive subscriptions and usage-based APIs. A strong open-weight model pressures that model. It tells developers: “You do not have to rent intelligence forever. You can own more of the stack.”
That is why this could sting.
MiniMax Is Not Playing Alone
MiniMax sits inside a brutally competitive Chinese AI market.
Reports name DeepSeek, Zhipu, and Moonshot AI as key rivals. The Economic Times also mentions other Chinese players passing the trillion-parameter threshold, with Meituan’s LongCat-2.0 and DeepSeek’s V4-Pro described as having around 1.6 trillion total parameters.
That context matters. MiniMax is not launching this model into a quiet pond. It is jumping into a koi pond where every fish has a rocket engine.
Chinese AI companies have been racing to release capable models at aggressive prices. Some have used open-source or open-weight strategies to gain developer adoption quickly. DeepSeek, in particular, helped prove that Chinese labs could produce models that attract global developer attention without copying the exact business model of US AI giants.
MiniMax appears to be making a similar bet, but with a giant scale twist.
If M3 Pro performs well, the company could stand out in a market crowded with capable labs. If it disappoints, the parameter count will become a punchline. Big numbers impress investors for about five minutes. Developers care about speed, cost, reliability, context length, tool use, coding strength, multilingual performance, and whether the model melts their infrastructure budget like butter on a skillet.
The benchmark gods will judge.
Why China’s Open-Model Push Is Working
Chinese open models have gained traction because they often offer a powerful combination: decent performance, lower cost, and flexible deployment.
That is catnip for developers.
Reports say Chinese open-source-based models are gaining adoption among developers looking for cheaper systems for high-volume, less critical tasks. The Economic Times reported that Chinese models from providers such as MiniMax, Z.ai, and DeepSeek are increasingly seen as lower-cost alternatives to proprietary US systems.
That phrase “less critical tasks” deserves attention. Not every AI workload needs the most expensive frontier model on Earth. A company may not need a $20-per-million-token genius to summarize support tickets, draft routine replies, classify documents, clean spreadsheets, or generate first-pass marketing copy.
Good enough becomes dangerous when it is cheap.
That is the commercial threat. If open Chinese models keep improving, they can eat the bottom and middle of the market. Then closed US labs must justify premium pricing for only the hardest tasks.
This is how disruption works. It rarely starts by beating the incumbent everywhere. It starts by becoming “good enough” for a large chunk of real work. Then it improves. Then the incumbent starts sweating through the blazer.
The Business Stakes for MiniMax
MiniMax is not just making a technical statement. It is making a market statement.
The Economic Times reported that MiniMax was founded in 2021, raised HK$4.8 billion, or about $614 million, in its Hong Kong initial public offering in January, and is planning a second listing on Shanghai’s STAR Market.
That gives the M3 Pro project a financial backdrop. A high-profile open-weight model could boost MiniMax’s reputation, developer ecosystem, and investor narrative. In AI, momentum matters. Perception matters. Leaderboards matter. GitHub stars matter. Adoption matters. The whole circus matters.
MiniMax also reportedly plans to launch H3, a frontier-level multimodal video generation model, later in July 2026. That points to a broader strategy. The company does not want to be seen only as a chatbot maker. It wants to compete across text, reasoning, multimodal generation, and possibly agentic systems.
That is ambitious.
But ambition cuts both ways. A 2.7-trillion-parameter model carries heavy training and serving costs. Even with efficient architecture, MiniMax will need strong infrastructure, optimization, and developer support. Open weights can drive adoption, but they do not automatically create revenue.
The model may be free to download. The business pressure will not be free.
The Mixture-of-Experts Clue
A model this large raises an obvious question: how do you run it without needing a small power plant and a dragon hoard of GPUs?
The likely answer is Mixture of Experts, or MoE.
The Economic Times report says the model’s architectural expansion relies heavily on Mixture of Experts engineering. MoE systems split a model into specialized subnetworks, often called experts. Instead of activating the entire model for every query, the system activates only a fraction of its total capacity.
That trick matters.
A dense 2.7-trillion-parameter model would be brutally expensive to run. But an MoE model can advertise a huge total parameter count while using only part of the model during inference. In plain English: the whole library exists, but the system only pulls the books it needs.
This can improve efficiency. It can also create specialized capability. One expert may help with coding. Another may help with math. Another may handle language tasks. The routing system decides what to activate.
Of course, MoE is not magic. Routing can fail. Training can get messy. Experts can become unbalanced. Serving infrastructure can become complicated. But if done well, MoE gives labs a way to build very large systems without paying the full cost of activating every parameter every time.
That is probably central to MiniMax’s plan.
Why Bigger Models Chase Harder Tasks
Standard chatbots handle short tasks well. They summarize, rewrite, translate, brainstorm, and answer basic questions. Useful? Absolutely. Revolutionary? Sometimes. Reliable enough for long, independent work? Not always.
The next AI frontier is harder.
Companies want systems that can plan a project, use tools, inspect files, write code, debug errors, reason across long documents, perform research, and complete multi-step workflows with less babysitting. That is the territory where bigger models, better architectures, and stronger training methods may help.
The Economic Times report says demand for autonomous systems capable of complex reasoning is driving the global push toward trillion-parameter AI models. It also notes that standard generative chatbots can struggle with independent, long-horizon decision-making.
This is the key strategic point.
The AI market no longer revolves around “Can it write a poem about a toaster?” We passed that station. The new question is: can it do useful work for an hour without wandering into the digital bushes?
If MiniMax’s model improves multi-step reasoning, coding, and agent-style tasks, it could become a serious tool for developers and enterprises. If it only inflates the parameter count without practical gains, users will shrug and go back to whatever works.
Performance, not size, wins the argument.
The US Labs Should Pay Attention
This MiniMax report lands at an awkward time for US AI labs.
American companies have spent huge sums training frontier models. Many then charge users through subscriptions, enterprise contracts, and APIs. That business model works best when the closed models remain clearly better than the open alternatives.
But if open-weight models get close enough, the pricing pressure becomes ugly.
TNW framed MiniMax’s reported model as another squeeze on US labs’ margins, arguing that cheap, capable open weights undercut the logic of charging a premium for frontier systems.
That analysis is directionally right.
Closed models can still win. They may offer better reliability, stronger safety systems, smoother tools, richer integrations, higher uptime, and top-tier performance. Enterprises also pay for support, compliance, and convenience. Nobody wants to debug a giant model cluster at 3 a.m. unless they enjoy pain as a hobby.
Still, open-weight models create leverage for buyers. They give companies a fallback option. They lower switching costs. They make premium vendors explain why their prices make sense.
That is healthy competition. Also terrifying competition, depending on which side of the invoice you sit on.
The Policy Shadow Over Open AI
There is one big complication: regulation.
Reports suggest Chinese authorities may want tighter controls on future releases of powerful open models. The Decoder notes that recent reports point to possible Chinese government efforts to tighten controls on future releases of such models. TNW also reported that Beijing has weighed curbs on who abroad can use its best models.
That creates uncertainty.
Open-weight AI can spread quickly across borders. Governments notice that. A powerful model can support business productivity and research, but it can also raise concerns around cybersecurity, surveillance, misinformation, military applications, and strategic competition.
So MiniMax’s reported plan sits inside a larger tension. China wants global AI influence. Chinese startups want developer adoption. Open models help with both. But regulators may not want the most capable systems flowing freely everywhere.
That tension could shape the final release. MiniMax might release full weights. It might release a restricted version. It might delay. It might face licensing limits. The public reporting does not yet settle those details.
So the honest answer is: the plan is exciting, but the final form remains uncertain.
In AI, “coming soon” can mean “next quarter,” “next year,” or “please stop asking our PR team.”
What Developers Will Watch First
When M3 Pro arrives, developers will not care about the press release for long. They will test it.
They will compare it against DeepSeek, Zhipu, Moonshot, Qwen-style models, US open models, and proprietary systems. They will look at reasoning benchmarks. They will test coding. They will check hallucination rates. They will measure latency. They will price out inference. They will see whether it works in English, Chinese, and multilingual settings.
They will also ask practical questions.
Can it run on available hardware? Can smaller teams host it? Does the license allow commercial use? Are the weights complete? Is the tokenizer painful? Does the documentation make sense? Does fine-tuning work cleanly? Does it support tool calling well? Can it handle long context? Does it collapse under real workloads?
These questions matter more than the headline number.
A model with 2.7 trillion parameters can still lose if it is expensive, slow, hard to deploy, or awkward to customize. Developers reward usefulness. They do not hand out trophies for “largest downloadable headache.”
Still, MiniMax has a strong hook. If it delivers a genuinely capable open-weight model at this scale, developers will test it immediately. Curiosity alone will drive the first wave.
After that, the model must earn its place.
The Bigger Picture: Open AI Is Becoming the Main Event
MiniMax’s reported 2.7-trillion-parameter model is not just a company story. It is a sign of where AI competition is heading.
The old frontier race focused on closed labs building the smartest models behind locked doors. That race still matters. But the open-weight race now matters just as much. Maybe more for developers. Maybe more for global adoption. Maybe more for pricing pressure.
MiniMax appears to be betting that scale plus openness can punch through a crowded market. It wants to build a model big enough to impress researchers, useful enough to attract developers, and open enough to spread.
That is a bold bet.
It could work. It could flop. The difference will come down to execution, not parameter theater. The world does not need another giant model that looks great in a headline and wheezes in production. It needs models that reason better, cost less, deploy cleanly, and help people build things.
For now, MiniMax has grabbed attention. A 2.7-trillion-parameter open-weight model will do that. It is the AI equivalent of walking into a room with a neon sign and a marching band.
The real test comes when the weights drop.
Then the hype clock stops.
And the benchmarking knives come out.
Sources
- The Next Web – “MiniMax is building China’s biggest AI model yet, and plans to open-source it”
https://thenextweb.com/news/minimax-2-7-trillion-parameter-open-source-model - The Decoder – “Chinese AI startup MiniMax plans to open-source a 2.7 trillion parameter model later this year”
https://the-decoder.com/chinese-ai-startup-minimax-plans-to-open-source-a-2-7-trillion-parameter-model-later-this-year/ - PANews – “MiniMax Plans to Launch a 2.7 Trillion Parameter Large Model”
https://www.panews.io/articles/019f40e0-4ed9-7332-832b-b5e5464deb4a - The Economic Times (Reuters) – “China’s MiniMax plans to launch giant 2.7 trillion parameter model”
https://economictimes.indiatimes.com/tech/technology/chinas-minimax-plans-to-launch-giant-2-7-trillion-parameter-model/articleshow/132263031.cms?from=mdr