Hi there 👋

Lobtensor

Tensor-native version control for AI models.

---

What is Lobtensor?

Lobtensor (lob) is version control built for tensors. Git manages source code; Lobtensor manages model weights, tensor metadata and tensor transformations; and Tensormorph visualizes, inspects and edits them.

Git          manages source code
Lobtensor    manages tensors, model weights, tensor metadata, transformations
Tensormorph  visualizes, edits and understands them

A Git user should be able to manage the history of a multi-terabyte model with Git's conceptual simplicity, without the system pretending tensors are just large files. The verbs are the same. The atom is different: the unit of change is the tensor and its 8 MiB chunks, not the file.

How it differs from Git and Git LFS

	Git / Git LFS	Lobtensor
Unit of change	a file	a tensor, split into 8 MiB chunks
Identity	hash of file bytes	BLAKE3 over canonical CBOR, in a content-addressed Tensor Merkle DAG
Diff	bytes or "binary files differ"	numerical: max|Δ|, cosine, tolerance-aware
Cost of a commit	the whole changed file	only the chunks that changed
Round-trip	byte copy of the file	every tensor's canonical bytes reproduced bit-for-bit
Compute	—	CPU is always complete; GPUs (CUDA, wgpu, Metal) only accelerate

A first session

lob init
lob import models/model.safetensors
lob status
lob commit -m "import checkpoint"
lob log
lob export HEAD --format safetensors -o out/
lob fsck

out/model.safetensors reproduces every tensor bit-for-bit. Beyond this path, lob has branches and merge, numerical diff, tensor deltas, remotes (file://, S3 layout and its own lob:// protocol), SSH signing, policy checks, Git pinning and hooks, and a local API server.

Principles

• Bit-exact or it is a bug. Export reproduces every tensor's canonical bytes.
• Stable identity. Anything that decides an object ID is specified, independently written and pinned by golden vectors.
• Cost scales with change. Fine-tune one layer, store one layer's chunks.
• CPU complete, GPU optional. Every operation runs on CPU; accelerators only make it faster.
• One container format first. SafeTensors (single, sharded and indexed) is the tracked format today; other adapters arrive by design decision, not by accident.

Status

Working, pre-1.0. lob round-trips real SafeTensors checkpoints bit-exact, and the on-disk format freeze (v1) is the next milestone. Expect formats to change until then.

Repositories

Repository	What it is
lobcli	the lob command-line tool and its Rust crates

_{Part of the Tensormorph ecosystem · Apache-2.0}