BlazeX: A Deterministic Archiving Format and Efficient Distribution Solution for Large Language Model Weights

BlazeX is an archiving format designed specifically for large language models (LLMs). It compresses model update sizes to 38-48% of full downloads using binary differential patch technology while ensuring byte-for-byte exact restoration. It addresses resource waste issues in current LLM distribution, such as the lack of standardized packaging and the need for full re-downloads for updates, and provides a complete archiving and differential solution.

Current LLM distribution lacks a standardized packaging layer. A 70B parameter model exists as a directory containing multiple .safetensors shards; each update (base version or fine-tuned variant) requires a full re-download; there are no standard community tools to answer questions about differences between models. This situation causes resource waste including duplicate storage, redundant bandwidth, and long download waiting times for users. BlazeX proposes solutions to these pain points.

BlazeX adopts a concise and stable format design with the structure [MAGIC 8B][VERSION4B][HEADER_LEN8B][HEADER JSON][RAW TENSOR DATA...]. Its features include:

• Header is human-readable using JSON;
• Tensor data is stored in original little-endian byte order while preserving the original type;
• Versioned management ensures backward compatibility;
• No built-in compression—users can choose external tools (e.g., zstd, lz4).

BlazeX generates patches by comparing tensors one by one, identifying changes using xxh3 checksums, and applying XOR + zstd incremental encoding (SplitStream for F16/BF16, sparse XOR for sparse integer tensors). This technology is its core value, significantly reducing update sizes.

Actual tests show:

• Qwen2.5-7B base to instruction fine-tuned patch: 6.1GB (40.1% savings);
• Qwen2.5-14B patch: 11.3GB (38.3% savings);
• Llama3.1-8B patch:7.7GB (47.9% savings). Compression efficiency improves as model size increases. Patch application ensures byte-for-byte exact restoration, verified via xxh3 checksums. Tests cover multiple model tensors, all passing SHA-256 integrity checks with no quality loss.

BlazeX provides a complete command-line toolchain:

• Packaging: Convert HuggingFace directory to .blz file;
• Information query: Metadata, tensor list;
• Extraction: Selectively export specific layers;
• Export: Convert to SafeTensors, PyTorch binary, or GGUF v3 (supports real-time quantization such as F16/BF16 downsampling, Q8_0, etc.). Its differential codec operates on raw BF16 weight bytes without architecture-specific assumptions, supporting cross-ecosystem use.

For developers: Efficient distribution of fine-tuned versions; For hosting platforms: Reduced storage costs; For users: Shorter update waiting times. BlazeX fills a gap in the LLM ecosystem, providing a solution for efficient distribution and version management. As model sizes grow and the fine-tuning ecosystem flourishes, its importance will become more prominent.