DeepSeek V4 Open Source Imminent! GitHub Codebase Reveals 'Model1' as the Star of the Spring Festival

On January 21, 2026, the tech community discovered a key clue in DeepSeek's FlashMLA optimization library—a new model identifier named "Model1" appeared in recent code commits. These commits span multiple files and mention Model1 alongside the current V3.2 model, clearly indicating that this is a separate new branch.

This discovery coincides with an exclusive report from tech media outlet The Information earlier this month—DeepSeek is intensely preparing to launch its next-generation flagship model, codenamed V4, during the Lunar New Year in mid-March. From the clues in the open source community to the cross-verification by authoritative media, the speculation that Model1 is V4 is almost a foregone conclusion.

A Technical Tribute for the First Anniversary of Open Source

Exactly one year ago, the release of DeepSeek-R1 opened a new era of open source LLMs and became the most liked model on the Hugging Face platform. Now, on the first anniversary of R1's open sourcing, the code for V4 has quietly appeared, continuing DeepSeek's commitment to the open source field.

Unlike R1, which focused on reasoning capabilities, V4 is expected to bring comprehensive architectural upgrades. According to leaked information, V4 has a total of 671 billion parameters, adopting a Mixture of Experts (MoE) architecture and Manifold Constrained Hyperconnection (mHC) technology, with the addition of the Engram memory module potentially being its biggest highlight.

The tech community is paying close attention to updates in the DeepSeek FlashMLA repository on GitHub, as every commit could reveal more technical details about V4. From the current commit records, DeepSeek engineers are intensively optimizing sparse attention mechanisms and FP8 inference support, all of which are preparations for the large-scale deployment of V4.

Engram: A Game-Changing "Memory Plug-in"

Most anticipated is the potential integration of Engram technology in V4, a revolutionary architectural innovation recently open-sourced by DeepSeek. Its core concept is to separate "memory" from "computation"—traditional Transformer models need to use multiple layers of networks to recognize fixed phrases and common sense, while Engram allows the model to directly retrieve static knowledge through a lookup-table "super dictionary", freeing up deep networks to focus on complex reasoning.

This design brings counter-intuitive results: even for tasks requiring strong reasoning like mathematics and code generation, performance is significantly improved. Because the model no longer needs to distract its computing power for "rote memorization." Research shows that Engram-27B improves by 5.0 points on BigBench Hard reasoning tasks and 3.0 points on HumanEval code generation compared to a pure MoE baseline.

More importantly, the Engram architecture supports storing a "knowledge base" of hundreds of billions of parameters in cheap CPU memory, with less than 3% throughput loss during inference. This means that the cost of enterprise deployment of large models is expected to be significantly reduced, no longer relying entirely on expensive GPU VRAM.

From Community Buzz to Industry Expectations

Discussion about the Model1 code leak has been going on for several days on Reddit's MachineLearning section and Hacker News. Developers are analyzing the latest commits on GitHub line by line, trying to piece together the full picture of V4 from code changes.

"This is typical DeepSeek style—no excessive marketing teasers, let the code speak for itself," commented a senior developer on a tech forum. This "speaking with technology" approach has instead accumulated stronger trust and expectation in the tech community.

From the open source success of R1 to the momentum of V4, DeepSeek is establishing a new open source paradigm: before the official release, through progressive updates of the codebase, allow the community to participate and adapt in advance. This strategy not only lowers the adoption threshold for new models but also creates a unique atmosphere of community co-construction.

Replicating the Successful Strategy of "Spring Festival Launch"

DeepSeek seems to be well versed in the golden rule of "Spring Festival Launch." Last year, R1 was released during the Lunar New Year, taking advantage of the relatively free time of global developers to quickly form viral spread in the open source community.

This timing choice is full of strategic considerations: avoiding the dense release schedule of Western tech companies and occupying a dominant position in the relatively quiet tech news cycle. At the same time, developers and researchers have more leisure time during the Spring Festival and are more willing to deeply test and explore new models, thus forming deeper word-of-mouth spread in the community.

With the continuous updates of Model1 code, readers who want to get the latest news and deep interpretation of DeepSeek V4 can continue to follow the tracking reports on deepseekv4.app. This independent tracking website has organized the full chain of information from code leaks to technical analysis, becoming the preferred channel for many developers to obtain V4 progress.

From R1 to V4, DeepSeek has always been proving one point: true innovation lies not only in model scale but also in the perfect combination of architectural wisdom and open source spirit. When the code of Model1 spreads quietly on GitHub, what we see is not only the birth of a new model but also the opening of a new paradigm. This Lunar New Year, the open source AI world may welcome another major breakthrough.