MiniMax M3 Pro: What We Know About the 2.7 Trillion-Parameter Open-Weight AI Model

MiniMax M3 Pro vs. Prior MiniMax Models

MiniMax's model family has grown rapidly in just over one year. The following table shows how the planned M3 Pro compares to its predecessors:

The jump from M3's ~428B total parameters to 2.7 trillion is a ~6.3x increase. If MiniMax continues its Mixture-of-Experts (MoE) approach, the active parameter count per forward pass could remain manageable—perhaps in the tens to hundreds of billions—while the total parameter footprint becomes unprecedented for an open-weight release ^MR.

How It Stacks Up Against DeepSeek V4 Pro

DeepSeek V4 Pro, released April 23, 2026, is the current open-weight benchmark leader in pure coding performance. Here is how the two compare, and where the planned 2.7T model would fit:

• DeepSeek V4 Pro: 1.6T total parameters, 49B activated per token, MIT-licensed. It achieves the highest Codeforces rating of any language model (3206) and 80.6% on SWE-bench Verified—best among open-weight models ^MAA. It is a pure-text algorithmic specialist, strongest on coding throughput and reasoning.
• MiniMax M3 (current): ~428B total / ~23B active. Stronger on multimodal tasks and mathematics (averaging 85.7 vs 31.7 on one aggregate), but weaker on pure coding benchmarks than DeepSeek V4 Pro ^B. It specializes in multimodal agentic coding, long-context work, and computer-use tasks.
• Planned 2.7T model: Would dwarf both in total parameters. As one analysis notes, "they barely compete"—M3 targets multimodal agentic tasks while V4 Pro optimizes for pure-text algorithmic work ^G. A 2.7T model could potentially bridge or redefine these categories, though architecture details are not yet public.

Crucially, the new model's name may differ at launch. While internal sources refer to "M3 Pro," Reuters' report does not use that codename, and the final name is unconfirmed ^RFF.

The Broader Chinese Open-Weight Model Wave

The M3 Pro is the leading edge of a sustained push by Chinese AI labs to release powerful open-weight models. This wave includes:

• DeepSeek V4 Pro (April 23, 2026): 1.6T MoE, MIT license, frontier coding performance ^MA
• Tencent Hy3 (Hunyuan 3.0): Preview released April 23, 2026; full open-weight release July 6, 2026 under Apache 2.0. A 295B-parameter MoE model (21B active, 192 experts) with 256K context, led by ex-OpenAI researcher Shunyu Yao ^BJD
• MiniMax-M1 (June 2025): 456B total parameters, billed as the world's first open-weight large-scale hybrid-attention reasoning model, Apache 2.0 ^BV
• MiniMax-M3 (June 2026): ~428B total, frontier multimodal coding ^W
• Ant Group Ling-2.6-flash (April 22, 2026): 104B/7.4B active efficiency-focused MoE ^L

This reflects a deliberate strategy: Chinese labs release open-weight models under permissive licenses (Apache 2.0, MIT) that allow commercial use and modification. This approach bypasses export controls on hardware by making the models themselves freely available to global developers, and it positions openness as a competitive advantage against Western labs like OpenAI, Anthropic, and Google that keep frontier models closed ^GRV.

What This Means for Competition with Western AI Labs

The planned 2.7T model, if realized, would make several statements:

• Scale race: It would surpass any known open-weight Western model in total parameters, challenging the narrative that US labs hold an unassailable compute advantage ^R
• Benchmark parity: MiniMax M3 already edges GPT-5.5's 58.6% on SWE-bench Pro with 59.0%, while DeepSeek V4 Pro's 80.6% on SWE-bench Verified is the best among any open-weight model ^CA
• Price competition: DeepSeek V4 Pro charges ~$0.435/M input tokens vs MiniMax M3 at ~$0.30/M, both well below GPT-5.5 pricing on API providers ^O
• Talent flow: Tencent's Hy3 was led by a hire from OpenAI, illustrating that AI talent moves in both directions ^B

The trend signals that China is no longer just a hardware-constrained follower—it is a leader in open-weight model releases, using openness as a competitive moat against Western closed-model leaders ^GR.

The Sheer Computational Challenge of a 2.7T Model

Deploying a 2.7 trillion-parameter model at scale poses extreme challenges, even before considering training costs:

• Memory requirements: In 8-bit quantization, loading 2.7T parameters requires ~2.7 TB of GPU memory. In FP16, that rises to ~5.4 TB. Even with a cluster of 80 GB H100s, you would need 34 GPUs just to hold 8-bit weights—and far more for inference, which also requires KV cache and activation memory. A practical deployment would require hundreds of GPUs.
• MoE efficiency: MiniMax's M1 and M3 both use MoE, activating only a fraction of total parameters per token (e.g., 23B out of 428B for M3) ^MW. If the 2.7T model follows a similar active-parameter ratio (e.g., ~135B active), inference could be feasible on a smaller cluster than the total parameter count suggests—but still far beyond a single node.
• Inference infrastructure: Serving such a model at scale demands sophisticated tensor parallelism, pipeline parallelism, and expert parallelism across many nodes with high-bandwidth interconnects.
• Training cost: Training a 2.7T model would require tens of thousands of GPU-months, even with efficient MoE training techniques. However, MiniMax has shown strong efficiency: M1 was trained with 70% less compute than DeepSeek R1 while achieving competitive benchmarks ^CV.
• Distribution: Simply hosting 5+ TB of model weights on platforms like Hugging Face is a logistical challenge.

That said, MiniMax's M3 already uses sparse attention mechanisms to achieve a 1M-token context window efficiently ^W. The company may employ similar architectural innovations for the 2.7T model, potentially making deployment less daunting than the raw parameter count suggests.

Growing Global Adoption of Chinese Open-Weight Models

Chinese open-weight models have gained significant traction globally for several practical reasons:

• Pricing: DeepSeek V4 Pro charges ~$0.435/M input tokens while MiniMax M3 is at ~$0.30/M—both far cheaper than most Western APIs ^O
• Permissive licensing: Apache 2.0 (MiniMax M1, Tencent Hy3) and MIT (DeepSeek V4 Pro) allow commercial use, modification, and redistribution without restrictions ^MD
• Benchmark competitiveness: These models match or exceed GPT-5.5 and Claude on key benchmarks including coding, reasoning, and mathematics ^CAA
• Developer ecosystem: Weights are hosted on Hugging Face and integrated into platforms like OpenRouter and Kilo Code, making them immediately accessible to global developers ^WK

The M3 Pro, if it ships on schedule and as planned, would be the strongest statement yet in this direction: a Chinese open-weight model that is not merely competitive on benchmarks or cost, but the largest model of any kind, open or closed, that the world has ever seen.

Key Takeaways

• MiniMax is building a 2.7 trillion-parameter open-weight AI model, internally codenamed M3 Pro, with a possible Q3 2026 release ^RT
• The model would be 6.3x larger than MiniMax's current M3 in total parameters, surpassing DeepSeek V4 Pro (1.6T) and any known open-weight model ^MMR
• It is part of a broader wave of Chinese open-weight releases including DeepSeek V4 Pro, Tencent Hy3, and earlier MiniMax models, all under permissive licenses ^GRB
• Deployment challenges are extreme but not insurmountable, especially if MiniMax continues its MoE approach and efficiency engineering ^MC
• The naming is not final: internal codename "M3 Pro" may change before official release ^RF