In January 2025, DeepSeek-R1 was launched. At this time, on the first anniversary of the release of the R1 model, DeepSeek's new model "MODEL1" was exposed. On January 21, Beijing time, DeepSeek's official GitHub repository updated a series of FlashMLA codes. With the help of AI, the total number of code files: 114 (including .py, .md, .txt, .sh, .cpp, .cu, .h files) was analyzed, and a previously undisclosed model architecture identifier "MODEL1" was discovered, which was mentioned 31 times in total.
FlashMLA is DeepSeek's original software tool that is deeply optimized for NVIDIA Hopper architecture GPUs and is specifically designed to accelerate the "inference generation" of large models. The implementation of this algorithm is based on MLA (multi-layer attention mechanism), which is one of the key technologies for DeepSeek models (such as V2 and V3) to achieve low cost and high performance. It is used to reduce memory usage at the model architecture level and maximize the use of GPU hardware.
MODEL1 is one of the two main model architectures supported in DeepSeek FlashMLA, the other being DeepSeek-V3.2. It is speculated thatMODEL1 is likely to be an efficient inference model. Compared with V3.2, it has lower memory usage and is suitable for edge devices or cost-sensitive scenarios. It could also be a long sequence expert, optimized for 16K+ sequences, suitable for long context tasks such as document understanding and code analysis. It may also be a long sequence specialist, optimized for 16K+ sequences,Suitable for long context tasks such as document understanding and code analysis.
In addition, the hardware implementation of MODEL1 spans multiple GPU architectures. There are two versions on NVIDIA H100/H200 (SM90 architecture): model1_persistent_h64.cu is used for 64-head configuration, and model1_persistent_h128.cu is used for 128-head configuration. There is a dedicated Head64 kernel implementation on the latest B200 (SM100 architecture), and the Head128 implementation of SM100 only supports MODEL1 and does not support V3.2.Some people speculate that DeepSeek has specially optimized the architecture of MODEL1 to adapt to NVIDIA's new generation of GPUs..
The main models that DeepSeek has released are representatives of two technical lines: the V series "all-round assistant" that pursues the ultimate comprehensive performance and the R series "problem solver" that focuses on complex reasoning.
V3, launched in December 2024, is an important milestone for DeepSeek. Its efficient MoE architecture has established a strong comprehensive performance foundation. Since then, DeepSeek has rapidly iterated on V3, released V3.1 to strengthen reasoning and agent capabilities, and launched the latest official version V3.2 in December 2025. At the same time, a special version V3.2-Speciale focused on solving difficult mathematics and academic problems was also launched.
R1, released in January 2025, uses reinforcement learning to perform outstandingly on complex reasoning tasks such as solving mathematical problems and code programming, and pioneered the "deep thinking" mode.
Technology media The Information broke the news at the beginning of the month that,DeepSeek will launch a new generation of flagship AI model - DeepSeek V4 during the Lunar New Year in mid-February this year, which will have stronger coding capabilities.
Previously, the DeepSeek research team released two technical papers, respectively introducing a new training method called "Optimized Residual Connection (mHC)" and an "AI Memory Module (Engram)" inspired by biology. This move has caused users to speculate that the new model being developed by DeepSeek may integrate these latest research results.