Executable programs and extensions

TensorFold Living Weights and RAG store knowledge differently

TensorFold 1.0.8's Living Weights experiment writes learned facts into a Nemotron folder. Start only from a copy.

RAG¹ retrieves relevant passages from a document store and places them in the prompt². Living Weights writes facts into the model folder so later runs can use them. TensorFold 1.0.8 documents a Mac workflow for Nemotron 3.5 Lightning with --slide. Learning rewrites files and has no automatic undo, so create a separate copy, verify its path, preserve the original, and test only on the copy.

Start with how each method supplies information

RAG retrieves passages related to a question and adds them to the model input. When documents change, update the store; retrieved source text can be shown as evidence. The model weights remain unchanged.

Operational differences between RAG and Living Weights
Aspect	RAG	Living Weights
Storage	Document index	Weight files in the model folder
Source evidence	Retrieved passages can accompany an answer	Learned facts may affect replies
Update	Edit documents and refresh the index	Restore a backup copy to remove learned facts

Living Weights learns supplied facts into files for a supported model. Later requests may reflect them without retrieving the same text into the prompt, but this is different from a document search system that finds and cites source passages.

A dark monitor, a compact desktop computer, and a PC case with several visible fans — Point learning at a model copy kept apart from the original.

Stay within the model scope documented for 1.0.8

TensorFold 1.0.8 marks Living Weights as experimental and documents learning on Apple-silicon Macs and NVIDIA GB10. The Mac example here uses the Nemotron 3.5 Lightning MLX³ 4-bit checkpoint⁴ with --slide.

Do not infer that --slide works with other checkpoints. Qwen3.8-27B and GLM families are not listed as supported in this release guidance; check the README before using any model/platform pair that is not explicitly named.

Living Weights and qualified models in TensorFold 1.0.8

Preserve the original and learn on a separate copy

tensorfold pull downloads the model into the Hugging Face cache. List the snapshot directories as shown below, inspect them, and set src to the exact revision for the Nemotron checkpoint; do not pick an arbitrary first entry. Copy only when the new destination ~/models/nemotron-living does not exist. cp -cRL can clone on the same APFS volume; a different volume needs room for a full copy.

Choose a revision and make a fresh copy

tensorfold pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit
cache="$HOME/.cache/huggingface/hub/models--TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/snapshots"
ls -1d "$cache"/*/
# Inspect the listed revisions, then paste the exact chosen directory below.
src="$cache/replace-with-exact-revision-from-list"
dst="$HOME/models/nemotron-living"
test -d "$src" && test ! -e "$dst" || { echo 'Check source and ensure destination is new'; exit 1; }
mkdir -p "$HOME/models"
cp -cRL "$src" "$dst"
tensorfold serve "$dst" --slide --name nemotron-living

Learning rewrites files in the model folder, and the README says there is no automatic undo. Use pwd or ls to confirm the target is the copy rather than the original cache. Read the actual ID from /v1/models, send a baseline question with that ID, then compare the same ID and question after learning.

Check the model ID and baseline answer

curl -fsS http://127.0.0.1:8080/v1/models
# Set this to the exact id returned above; --name normally sets it to nemotron-living.
MODEL_ID='paste-the-id-returned-by-v1-models'
curl -fsS http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' -d "{\"model\":\"$MODEL_ID\",\"messages\":[{\"role\":\"user\",\"content\":\"What color do I like?\"}],\"max_tokens\":32,\"temperature\":0}"

Send one learning request

curl -N http://127.0.0.1:8080/v1/slide/learn -H 'Content-Type: application/json' -d '{"text":"I like blue."}'

Check the result with the same question after learning

MODEL_ID='paste-the-same-id-from-v1-models'
curl -fsS http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' -d "{\"model\":\"$MODEL_ID\",\"messages\":[{\"role\":\"user\",\"content\":\"What color do I like?\"}],\"max_tokens\":32,\"temperature\":0}"

Official copy-and-serve example

A desk with a large monitor showing code-like lines, a desktop tower, and a laptop — Write the fact clearly and make it easy to verify before teaching it.

Check what it remembers and restore if needed

Check the success log for the learning request separately from the model's answer. Compare the answer to What color do I like? with the baseline you recorded. If the answer is wrong or unwanted, stop the server and replace the model folder with the preserved original or backup⁵ copy instead of trying to overwrite it through more learning.

Before adding another fact, repeat the same test question. Keep pre-learning and modified directories separate so the folder to restore is clear.

Learning target folder and rollback condition

Printed reports and an open notebook around a laptop displaying documents and charts — After learning, repeat the same question and compare against the original.

Terminology notes

Retrieval-Augmented Generation — A method that retrieves material relevant to a query, adds it to the model input, and generates an answer from that context.
Back to the text
Prompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
MLX — Apple’s machine-learning framework, which runs models using memory shared by the CPU and GPU on Apple Silicon.
Back to the text
Checkpoint — A file containing saved model weights and related state. Versions or tasks in one model family may use different checkpoints.
Back to the text
Backup — Saving settings and data to a file so the same setup can be returned to after a hardware failure.
Back to the text

Frequently asked questions

What should you check first in this article?: RAG retrieves relevant passages from a document store and places them in the prompt. Living Weights writes facts into the model folder so later runs can use them. TensorFold 1.0.8 documents a Mac workflow for Nemotron 3.5 Lightning with `--slide`. Learning rewrites files and has no automatic undo, so create a separate copy, verify its path, preserve the original, and test only on the copy.
When is it better not to choose this?: `tensorfold pull` downloads the model into the Hugging Face cache. List the snapshot directories as shown below, inspect them, and set `src` to the exact revision for the Nemotron checkpoint; do not pick an arbitrary first entry. Copy only when the new destination `~/models/nemotron-living` does not exist. `cp -cRL` can clone on the same APFS volume; a different volume needs room for a full copy.