One year after the groundbreaking release of DeepSeek-R1, Chinese AI company DeepSeek appears to be preparing its next major model launch. References to a previously unknown system called "Model 1" have emerged in recent code updates to the company's GitHub repository, fueling speculation that this could be the rumored DeepSeek-V4 set for release around the Chinese New Year.

The discovery comes exactly one year after DeepSeek released its R1 model on January 20, 2025, which became the most-liked model on Hugging Face's platform and marked what the company called "the DeepSeek moment" in open-source large language models.

Multiple code commits to DeepSeek's FlashMLA repository over recent days contain explicit references to "Model1," including dedicated configuration files that place it alongside the existing DeepSeek-V3.2. The frequency and prominence of these references suggest active development of a major new system rather than incremental updates to existing models.

The technical specifications visible in the code indicate significant architectural changes, including a shift to 512-dimensional parameters and optimization for NVIDIA's next-generation Blackwell GPU architecture, suggesting DeepSeek is preparing a substantial leap in capabilities.

Architecture Shift Signals New Model Generation

The most telling evidence that Model1 represents a new generation rather than an iteration appears in the repository's core architecture files. Code in the csrc/api/common.h file shows Model1 configured with 512-dimensional head parameters, a departure from the 576-dimensional configuration used in DeepSeek-V3.2.

This architectural change suggests DeepSeek has redesigned its Multi-head Latent Attention (MLA) structure. The V3 series used an asymmetric design combining 128-dimensional RoPE with 448-dimensional latent dimensions. The shift to a standardized 512-dimensional configuration may optimize for hardware alignment or represent improvements in latent compression ratios.

The code repository's dispatch logic treats Model1 as a distinct branch parallel to V3.2 rather than as a derivative, indicating a fundamental architectural departure. According to DeepSeek's historical naming conventions, a flagship architecture transition following V3.2 would logically be designated V4.

Blackwell GPU Optimization Points to Performance Focus

The code updates reveal extensive optimization work for NVIDIA's Blackwell architecture, suggesting DeepSeek is positioning Model1 for next-generation hardware. New interfaces include FMHACutlassSM100FwdRun specifically targeting Blackwell's instruction set, with documentation noting that operation on B200 GPUs requires CUDA 12.9.

Performance metrics embedded in the code indicate that even without full optimization, Sparse MLA operators achieve 350 teraflops on B200 hardware. On current-generation H800 GPUs with SM90a architecture, Dense MLA operations reach 660 teraflops throughput.

The Blackwell-specific optimizations suggest DeepSeek is preparing Model1 for deployment on infrastructure that may not be widely available until later in 2026, indicating a strategic timeline aligned with hardware availability rather than immediate release.

Token-Level Sparse MLA Introduces Efficiency Mechanism

Among the most significant technical additions visible in the code is the introduction of token-level Sparse MLA, a new computational approach not present in previous DeepSeek models. Test scripts now include both test_flash_mla_sparse_decoding.py and test_flash_mla_dense_decoding.py, indicating parallel sparse and dense processing capabilities.

The sparse implementation uses FP8 precision for key-value cache storage while performing matrix multiplication in bfloat16 precision to maintain accuracy. This hybrid approach suggests Model1 is designed to handle extremely long context windows by reducing memory pressure through selective sparsification during inference.

The architectural addition of sparse processing represents a response to one of the key challenges in scaling language models: maintaining performance as context length increases while managing memory constraints. The dual-path design allows the system to dynamically balance between computational efficiency and precision requirements.

New Mechanisms Suggest Capability Expansion

While the code commits primarily address operator-level implementations, references within the dispatch logic point to additional new features. The repository structure suggests Model1 incorporates Value Vector Position Awareness (VVPA), potentially addressing position information degradation in long-text processing that affects traditional MLA architectures.

Code comments also reference an "Engram mechanism," though implementation details remain limited in the visible commits. Based on its placement within distributed processing modules, this feature likely relates to distributed storage optimization or advanced key-value compression to support Model1's high-throughput requirements.

The combination of architectural changes, hardware optimization, and new processing mechanisms visible in the code suggests Model1 represents a comprehensive redesign rather than incremental improvements. Whether this system will be officially designated as DeepSeek-V4 remains unconfirmed, but the technical evidence points to a major release in DeepSeek's development roadmap.