A language model is a file full of numbers. Those numbers decide everything it says. Neyra finds the ones behind a fact or a habit and changes them.
- On your machine
- Mac, Linux or Windows. No cloud, no account.
- No retraining
- You change one thing. You don’t teach the whole model again.
- Your original stays
- An edit is a separate small file. Delete it and nothing happened.
See it work
We told a model the capital of France is Rome.
$ neyra edit Qwen/Qwen3-0.6B "The capital of [France] is" RomeOne command. The brackets mark what the fact is about.
efficacy 0.0% -> 100.0%Before, the model preferred the real answer. Now it prefers Rome.
decoded effi 0.0% -> 100.0%Same when it writes the answer out itself, not only when we score it.
neutral text: KL 0.0006 nats/tokenEverything unrelated barely moved.
fluency 4.13 -> 3.97It still writes normally.
pass exact_revertThe edit undoes perfectly, down to the last bit.
FAIL specificity
not claimed: the fact took, but nothing
checked its neighbours.Neyra refuses to call this a success, because we gave it no nearby facts to test. Add some and it checks those too. If an edit does damage, it tells you.
Real output from Qwen3-0.6B, trimmed. This run took six minutes on a busy CPU with 4.5 GB of memory. On Apple Silicon our benchmark took 3.2 seconds per edit.
What it does
Twelve things you can do to a model.
Change a model
- Rewrite a factGive it a sentence and a new answer. The model says the new answer from then on.
neyra edit MODEL "The capital of [France] is" Rome - Stop a bad habitTest a model for loops and runaway thinking, then write a patch that targets the loop.
neyra probe MODEL
neyra fix MODEL -o fix.neyra - Combine editsMerge several edits into one, with TIES, DARE, SLERP and others. The order never changes the result.
neyra merge MODEL a.neyra b.neyra -o both.neyra - Make the edited modelBuild a normal model from your original plus the edit, ready to run anywhere.
neyra apply MODEL edit.neyra -o edited-model
Understand a model
- See what someone changedCompare two models and see which parts moved and how, often from a few percent of the download.
neyra diff BASE OTHER --sample 4 - Find where your fine-tune wentIt worked in training and not in the server. This names the step that lost it: merge, convert or quantize.
neyra hop BASE ADAPTER MERGED model.gguf - Check a merge really happenedFinds adapters that were never applied, and matches PEFT and mlx-lm merges bit for bit.
neyra check BASE MERGED --adapter ADAPTER - Catch bad training data firstReads your dataset the way your trainer will and tells you what will go wrong, before you pay for the run.
neyra tune lint data.jsonl MODEL --trainer trl - Find out what a model really isSame weights, a known conversion, a real derivative, or unrelated to the base it claims.
neyra identity BASE MODEL
Use it with everything else
- Export to the tools you useTurn an edit into a LoRA for PEFT or llama.cpp. Read edits from Delta League, OpenWSP and others.
neyra export MODEL edit.neyra --format peft -o adapter - Let an AI do the editingAn MCP server lets an agent try edits under a fixed budget, against a test it can’t see.
neyra mcp - Prove what you changedSign a record of the change that other people can check without trusting you.
neyra attest BASE RESULT --key signer.key -o change
What you need
Less than you’d think.
How good is it
What we measured, and what we haven’t.
Measured
- 100%of 200 unseen facts took on Qwen3-0.6B, one edit at a time.
- 95%still held when the question was reworded.
- 0.70score on neighbouring facts, down from 0.745. Edits do cost a little.
- 0false alarms from
diffin 250 rounding noise trials. - 312protocol tests, passed by two separate verifiers in Python and Go.
Not yet
- Nothing tested above 1B parameters. Mostly Qwen3 and Qwen3.5.
- It rewrites facts and removes problems. It hasn’t been shown to make a model smarter.
- ROME did better on the 20 cases compared so far. The full comparison isn’t finished.
- Large batches of facts can damage nearby facts. Neyra flags it when they do.
- Edits go into dense layers only. Quantized files can be checked, not edited.
Questions
Things people ask first.
Is it free?
Yes. It’s open source under Apache 2.0.
How is this different from fine-tuning?
Fine-tuning trains on examples and nudges everything. Neyra changes one chosen thing and tells you exactly what moved.
Can I break my model?
Your original file isn’t touched. The edit lives in its own file, and Neyra checks it undoes exactly.
Will it make my model smarter?
No. It can change what a model says and remove problems like loops. Adding real ability hasn’t been shown.
Does it work with GGUF?
You edit safetensors models. Neyra can read GGUF files to check them, and export a LoRA that llama.cpp loads.
What can I try right now?
neyra demo runs in about a tenth of a second, downloads nothing and needs no GPU.