China’s New Open-Source AI Just Crushed DeepSeek! (7/27/2025)

This video describes a new large language model (LLM) called M1 released by the Chinese company Miniax. Key points about M1 from the video and transcript are:

M1 is a groundbreaking open-source LLM with an extraordinarily large processing capacity, capable of handling 1 million input tokens and generating up to 80,000 token responses. This is far beyond previous models like GPT-4 (which handles about 125k tokens input), Claude 4 Opus, and DeepSeek R1, which max out around 128k tokens.

The model architecture is innovative: It contains 456 billion parameters in total but uses a Mixture of Experts framework that activates only about 46 billion parameters per operation by loading 32 specialized subnetworks depending on the tokens processed. This reduces computational load dramatically.

M1 uses a proprietary "Lightning Attention" mechanism instead of traditional quadratic attention, giving it linear scaling with sequence length and enabling efficient handling of long texts without prohibitive resource use.

Training was completed in about three weeks using 512 Nvidia H800 GPUs at a rental cost of approximately $535,000, significantly cheaper than DeepSeek's reported $5-6 million and far below GPT-4's estimated over $100 million training cost due to architectural and training optimization innovations.

It is fully open source with permissive licensing allowing complete on-premises deployment and no usage or subscription fees, aiming to democratize advanced AI capabilities worldwide.

M1's strengths include extended context understanding (handling millions of tokens), competitive reasoning and coding performance, and flexible task adaptability. It slightly trails DeepSeek on some benchmarks but exceeds other open models.

The project recommends using VLLM as a backend for serving the model for efficient memory pooling and expert model handling, with other standard transformer libraries also supported.

The release represents a major leap for open-source AI in capability, efficiency, and accessibility, potentially shifting global AI power balance.

Regarding smaller quantized versions, the video transcript and description do not explicitly mention quantized or smaller lightweight versions of the M1 model being released yet, though given the scale of the model (hundreds of billions of parameters globally, 46 billion active), future quantized variants would be relevant for wider practical deployment.

In summary:

Name: M1 language model by Miniax

Size: 456 billion parameters total, 46 billion active per token operation

Unique features: 1 million token input, Lightning Attention (linear scaling), Mixture of Experts, very low computational training cost

Open source: Yes, fully open source with permissive licensing, free with no usage fees

Quantized smaller versions: Not explicitly mentioned yet

Deployment: Supports VLLM backend, on-premises use possible

This model marks a significant advancement in open LLMs enabling unprecedented extended-context language understanding and cost-effective training compared to previous giants like DeepSeek and GPT-4. It is already being integrated across Asia and poses a challenge to Western closed-model dominance in AI.