neyra 0.2.1

Neyra edits the weights of local language models.

Install · Usage · Commands · Design · Results · Docs · Spec

Neyra writes facts and behaviour changes into safetensors checkpoints. An edit is a patch: about 170 KB for one fact on Qwen3-0.6B, tied to the digest of its base, and reversible bit for bit. neyra edit uses a MEMIT style closed form solve across several MLP layers. neyra fix finds the direction behind a reasoning loop and writes a patch that projects it out of the columns that write it.

It also reads changes made with other tools. diff recovers the form of an edit from two checkpoints, check recomputes a LoRA merge with the rounding PEFT and mlx-lm use, and hop walks base, adapter, merged model and GGUF to the first one that differs.

neyra diff on the loop edit from Sparks of Weight-Space Programming, reading 4% of both checkpoints over HTTP range requests. The authors' notes give the same formula with a strength of 6.

Install

pip install neyra        # or: uv tool install neyra
neyra demo

Needs Python 3.10 or later. The base install handles patches, diffs and verification with numpy and cryptography. To run models, add neyra[mlx] on Apple Silicon or neyra[torch] elsewhere. neyra edit needs MLX. neyra doctor lists the backends you have.

neyra demo takes about 0.1 s and downloads nothing. It builds a toy model and a LoRA in memory, then runs check on a merge that dropped the adapter and on one that kept it.

Usage

Paths and Hugging Face repo names both work.

neyra probe MODEL                             # loops, runaway thinking, missing answers
neyra fix MODEL -o fix.neyra                  # candidate patch for a loop direction
neyra edit MODEL facts.json -o edit.neyra     # write facts, measure, sign
neyra apply MODEL edit.neyra -o edited-model  # rebuild, checking digests

neyra diff BASE OTHER --sample 4              # compare two checkpoints from a few percent of the bytes
neyra check BASE MERGED --adapter ADAPTER     # recompute a LoRA merge and compare bits
neyra hop BASE ADAPTER MERGED model.gguf      # first artifact that differs

Commands

Edit

Inspect

Sign

Automate

Design

A patch is a safetensors file, so loading one runs no code. It records the digest of the tensors it expects and the digest it produces, and apply stops on a mismatch. The weights root hashes every tensor and ignores sharding, so a resharded model keeps its root.

merge sums deltas as integers and rounds once, because bf16 addition does not associate. At an entry with value 1 - 2^-8, three patches that each add 2^-8 give 1.0 in pairs and 1 + 2^-7 together.

neyra edit solves in float64 and measures the rounded weights the patch produces. Its gates cover efficacy, neighbouring facts, drift on neutral text and fluency. The limits ship in the code, not in your run, and the receipt keeps a FAIL as well as a PASS.

Output separates what you can recompute from what was measured. Sampled reads report their coverage. identity calls a model derived against distances measured on unrelated sibling models. A signature tells you who signed, so a receipt also lists what you need to rerun the claim.

hop takes artifacts in the order a fine-tune passed through them:

Results

So far these edits rewrite facts and remove loops. None has been shown to add ability. ROME led edit on the 20 development cases compared, and the full comparison is unfinished. Nothing has been tested above 1B parameters. The limits page has the rest.

Documentation

Guides: edit, merge, bisect, check, hop, identity, signing, lab.

Reference: specification, protocol notes, vectors, accuracy, limits, prior work, changelog.

Status

Early. Formats can change until v1.

See CONTRIBUTING.md to work on it and SECURITY.md to report a vulnerability. If you write a verifier from the spec and vectors, open an issue. The Go one here shares an author with the Python one.

License

Apache 2.0