A new GitHub project, vixhal-baraiya/qwenFS, has been trending this week, and many are wondering what it’s all about. The project embeds a working filesystem inside the weights of a large language model, Qwen3.5-2B, allowing it to store and manage files without the need for external tools or databases.
Background And History
The project’s filesystem is implemented as 8 transformer layers inserted before Qwen’s 24 layers, with attention mechanisms storing file content and handling path lookups and search indexing. The gated DeltaNet state maintains version tables, path tables, and link graphs, while MLPs parse tool calls and emit forced tokens for exact filesystem operations. This innovative approach enables the model to save, read, organize, and link its own files through normal tool calls, with the model answering every call itself.
How It Works
The filesystem works by having the attention layers hold the files, with a write’s text staying in the copy layer’s attention log, keyed by file index, version, and offset. A read is an exact copy, token by token, and each write or append is a segment, a version of its own. The model also uses a decaying DeltaNet counter to track offsets, which are decoded exactly to 600 tokens per segment. The filesystem saves to a compact image, with paths, segments, and links stored as 3-byte token IDs, and a fresh session mounts the image and pages a segment’s text in only when the model emits a read acknowledgment.
The Numbers
The model runs at approximately 20 tokens per second on Apple silicon with 16GB of memory, and supports up to 128 files, paths of up to 8 tokens, and about 590 tokens per write or append. Search functionality finds exact case-sensitive whole words up to 3 tokens, and the model’s top-1 predictions remain unchanged at all 139 checked positions.
What It Means Now
For families and individuals interested in artificial intelligence and language models, this project represents a significant breakthrough in the field. The ability to store and manage files within a language model’s weights has the potential to enable new applications and use cases, such as more efficient data storage and retrieval. As the project continues to evolve, it will be exciting to see how it is used and what benefits it brings to users. The code is available on GitHub under the MIT license, and a checkpoint built with the project’s merge tool falls under Qwen’s Apache-2.0 license, making it accessible to developers and researchers. With its innovative approach and impressive results, vixhal-baraiya/qwenFS is definitely a project to watch in the world of AI and language models, and its impact will be closely followed by the infopulsetoday.com community.
The project’s creator has made the code available for anyone to use, with instructions on how to clone, build, and run the project, including a chat app that allows users to interact with the filesystem. As the project continues to gain attention, it will be interesting to see how it is used and what new applications are developed using this technology, with the next step being to explore its potential uses and limitations, and to see how it can be used to benefit people in their daily lives, for families and individuals looking to leverage the power of AI and language models.
Sources
Photo: Daniil Komov / Unsplash (Unsplash License)
