---
license: apache-2.0
base_model: unsloth/Qwen3.5-9B
pipeline_tag: text-generation
tags:
  - code
  - change-analysis
  - commit
  - gguf
  - on-device
---

# Knot Delta

**On-device model that explains what changed.** Give it a commit and Knot Delta
writes a grounded report: what moved, why, and the risk. A 9B model, served as a
single quantized GGUF that runs on your machine.

Part of the **Knot AI** family, alongside
[**Knot Scribe**](https://huggingface.co/pointbreaklab/knot-scribe) (which writes
your commit messages). Delta is the change-intelligence layer inside
[Knot](https://pointbreaklab.com/knot/).

A fine-tune of **Qwen3.5-9B**, merged and exported to a text-only
`Q4_K_M` GGUF (~5.2 GB) for `llama.cpp` / Ollama.

## Capability scorecard (honest)

Every number is measured against **objective ground truth** (real git history),
reported **beside the baseline a lazy model would score**. No aggregate
"intelligence score" is quoted. Frozen held-out set: **325 commits across 9
unseen real repos** (vite, vue, svelte, astro, ionic, gumroad, OpenDream,
radicle, vllm).

| capability | number | baseline | verdict |
|---|---|---|---|
| **Change-type** (intent read from code), macro-F1 | **0.409** | majority-class 0.032 | **real — 12.9× base rate** (n=235) |
| **Cited files that exist in the diff** | **100%** | — | never hallucinates a file; holds on unseen Go |
| **Reports fully grounded** | **98.5%** | — | 325 held-out commits (first-pass 88%) |
| **Exact-line precision** | **0.64** | 0.94 lenient (right region) | honest limit: file + region reliable, exact line ~64% |

**Multi-repo generalization** (exact-line precision, 4 unseen repos):

| repo | language | verified | line precision |
|---|---|---|---|
| flask | Python | 11/12 | 0.74 |
| gin | Go (unseen) | 12/12 | 0.70 |
| ripgrep | Rust | 12/12 | 0.52 |
| frozen-325 | JS/TS + | 320/325 | 0.64 |

**Honest read:** Delta never cites a file that isn't in the diff (100% across
every repo, including Go, which is absent from its training data), and it always
points at the right region. What it does *not* yet do reliably is pin the exact
line — that lands 52–74% of the time, lowest on Rust. It knows which file and
which part changed; the precise line is still improving. One capability,
measured straight, with the parts that don't yet work said out loud.

## Use it

**llama.cpp**

```bash
llama-server --model kd-delta-9b-Q4_K_M.gguf
```

**Ollama**

```bash
ollama run pointbreaklab/knot-delta
```

## Model

- **Base:** Qwen3.5-9B
- **Format:** text-only `Q4_K_M` GGUF, ~5.2 GB
- **Runs:** `llama.cpp` / Ollama, CPU or GPU, fully local
- **License:** Apache-2.0