All posts

MiniMax to open-source 2.7 trillion parameter AI model

Manaal KhanJuly 8, 2026 at 6:31 PM4 min read
MiniMax to open-source 2.7 trillion parameter AI model

Key Takeaways

MiniMax to open-source 2.7 trillion parameter AI model
Source: The Decoder
  • MiniMax's M3 Pro would be 6.3x larger than their current flagship and the largest open-source model released
  • Release expected as early as Q3 2026, though Chinese government may tighten controls on future releases
  • Chinese open-source models are gaining traction for high-volume, cost-sensitive AI workloads

MiniMax, one of China's fastest-growing AI startups, is building a 2.7 trillion parameter language model it plans to release as open source later this year. The Information reported the news Tuesday, citing two sources familiar with the company's plans. If released, M3 Pro would become the largest publicly available AI model by a wide margin.

The timing matters. Chinese open-source models have carved out real market share in 2026, particularly among teams running high-volume inference on cost-sensitive tasks. MiniMax competes directly with DeepSeek, Zhipu, and Moonshot AI for this developer base. A model this size could shift the competitive balance.

Advertisement

How big is 2.7 trillion parameters?

MiniMax's current flagship, M3, runs 428 billion parameters. The new model would be roughly 6.3 times larger. For context, GPT-4 is estimated at around 1.8 trillion parameters, though OpenAI has never confirmed the figure. Meta's Llama 3 tops out at 405 billion.

Parameter count does not map directly to capability. But larger models generally handle complex reasoning, multi-step instructions, and nuanced tasks better than smaller ones. They also cost more to run. The open-source release would let teams fine-tune and deploy M3 Pro on their own infrastructure, a significant consideration for companies wary of API dependency or data privacy.

What's driving Chinese open-source momentum?

Three factors explain why Chinese AI labs have bet heavily on open weights. First, open source builds developer ecosystems fast. Alibaba's Qwen family and DeepSeek's models have accumulated millions of downloads on Hugging Face by letting anyone run inference locally. Second, it sidesteps the API pricing war with OpenAI and Anthropic. Teams can run inference at GPU cost, not per-token markup. Third, open release creates pressure on closed competitors. It is harder to charge premium prices when a capable alternative is free to download.

MiniMax has taken a slightly different path than its peers. The company built its consumer business around Talkie, an AI companion app popular in the US and Southeast Asia. Enterprise and developer tooling came second. An open-source 2.7T model signals a pivot toward infrastructure, not just applications.

Advertisement

Regulatory uncertainty looms

Recent reports suggest Beijing wants tighter controls on future open-source AI releases. The concern, broadly stated, is that open weights let anyone, including foreign adversaries, replicate Chinese AI capabilities without restriction. No specific regulations have been announced, but the timing creates risk for MiniMax's Q3 target.

MiniMax is backed by Alibaba and Tencent, with an estimated valuation exceeding $600 million as of late 2024 funding rounds. That investor profile could insulate the company from regulatory friction, or make it a higher-profile target. The outcome is unclear.

What this means for teams evaluating open models

If M3 Pro ships on schedule, it raises the ceiling for what open-source models can do. Current best-in-class options, Llama 3 405B, Qwen 2.5, DeepSeek V2, all cluster in the 70B to 400B range. A 2.7T model operating at comparable efficiency would be a different class of tool.

The practical question is inference cost. A model this size requires significant GPU resources. Running it locally is not an option for most teams without high-end hardware. Cloud inference, whether self-hosted or through a managed service, will set the real accessibility bar.

ℹ️

Logicity's Take

MiniMax's 2.7T model is a capability statement, not a cost-efficiency play. Most production workloads will not need a model this large. But the release could reset expectations for what open weights can achieve on hard tasks: long-context reasoning, agentic chains, and multi-modal synthesis. If you are building agents or complex workflows, the M3 Pro release is worth tracking. For commodity tasks like summarization or classification, smaller models like DeepSeek V2 or Qwen 2.5 72B remain more practical choices with far lower inference costs.

Frequently Asked Questions

When will MiniMax release the 2.7 trillion parameter model?

Sources cited by The Information suggest a release as early as Q3 2026, though the timeline could shift depending on development progress and regulatory factors.

How does MiniMax M3 Pro compare to GPT-4?

M3 Pro's 2.7 trillion parameters would exceed GPT-4's estimated 1.8 trillion. However, parameter count alone does not determine model quality. Architecture, training data, and fine-tuning matter as much or more.

Will M3 Pro be free to use?

MiniMax plans to release it as open source, meaning the weights will be downloadable. Running inference still requires GPU infrastructure, which carries compute costs.

What are the main competitors to MiniMax?

In China's AI market, MiniMax competes with DeepSeek, Zhipu AI, Moonshot AI, and Alibaba's Qwen team. Globally, open-source competition includes Meta's Llama and Mistral.

Also Read
Why your AI-first announcement keeps backfiring

Context on how companies position AI capabilities versus actual delivery

ℹ️

Need Help Implementing This?

Logicity helps product teams evaluate and deploy open-source AI models for production workloads. Contact us to discuss your infrastructure requirements.

Source: The Decoder / Maximilian Schreiner

Advertisement
M

Manaal Khan

Tech & Innovation Writer

Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in our Editorial Policy.