How it works¶

Here is the whole path, from a model name to a smaller model.

 "llama3.2:1b"
      │
      ▼
 1. FIND      Ollama manifest  ──►  weights file (.gguf) + template + settings
      │
      ▼
 2. READ      header + metadata + tensor list   (the weights are not loaded)
      │
      ▼
 3. UNDERSTAND  layers, heads, vocabulary, chat template
      │
      ▼
 4. MEASURE   weight statistics          (optional: run the model = trace)
      │
      ▼
 5. ANALYZE   list of removable parts, each with size and risk  →  plan.json
      │
      ▼
 6. SLIM      write a new .gguf without those parts, then check it
      │
      ▼
 small.gguf + Modelfile  ──►  llama.cpp / ollama create

1. Find¶

Ollama keeps a small manifest file for each model. It says which big file holds the weights and which small files hold the chat template, system prompt and settings. NeuronDump also scans the llama.cpp, Hugging Face and LM Studio folders for .gguf files.

2. Read¶

A GGUF file starts with a table of contents: settings (llama.block_count = 16…) and a list of every weight matrix with its name, shape and position. NeuronDump reads only this table, so it is fast even for huge files.

3. Understand¶

Weight names follow a pattern. For example, blk.3.attn_q.weight is layer 3, attention, "query" matrix. From the names in each layer, NeuronDump knows what kind of layer it is:

Layer contains	Kind
`attn_q`, `attn_k`, `attn_v`	attention
`ssm_…`	linear attention (state-space)
`ffn_…_exps`	mixture of experts
`ffn_up`, `ffn_down`	normal feed-forward

It also decodes the vocabulary into readable text and parses the chat template.

4. Measure¶

stats reads the weights in small pieces and computes how strong each part is. trace runs the model on a sentence and records what every layer does:

LAYER  LOGIT LENS (what the model would answer here)
10     "olit" 0.04   "polit" 0.02
11     " Rome" 0.35  " capitals" 0.07   ← the answer appears here
15     " Rome" 0.84

5. Analyze¶

NeuronDump lists what could be removed:

Candidate	Example	Risk
unused tokens	248 reserved tokens never used	medium
vocabulary trim	keep only the tokens your texts need	high
duplicate output matrix	exact copy of the input embeddings	low
a whole layer	one that changes the text very little	high
16-bit weights	could be stored in 8 bits	low
vision part	not needed for text-only use	low

6. Slim¶

You choose candidates, and NeuronDump writes a new file. It renumbers tokens and layers, fixes every setting that depends on them, then re-reads the file to check it:

Size      1.23 GiB -> 987.47 MiB (21.6% smaller)
Validate  OK (147 tensors, vocab 570, block_count 16)

A Modelfile is written next to it, so ollama create works directly.

CLI or web¶

Both use the same engine. neurondump cli opens an interactive prompt; neurondump web --open opens the same features in your browser, on your own computer only.

Next: getting started.