README.md

---
license: apache-2.0
base_model:
  - Qwen/Qwen3-4B
  - Qwen/Qwen3-1.7B
pipeline_tag: text-generation
library_name: gguf
tags:
  - knaif
  - planning
  - function-calling
  - gguf
---

# knaif — natural language → validated action plans

Fine-tuned **Qwen3** models for [knaif](https://github.com/…): they turn a natural-language
request into a strict JSON **action plan** (`{"plan": [...]}`) that deterministic code then
validates, expands, confirms, and executes through skill packages. **The model only proposes
the plan** — it never runs anything itself.

These are SFT fine-tunes trained on the **ffmpeg** and **documents** skills.

## Models in this repo

| File | Base | Params | Quant | Size | Surface |
|------|------|--------|-------|------|---------|
| `knaif-qwen3-4b-v1-q4_k_m.gguf`   | Qwen3-4B   | 4B   | Q4_K_M | 2.5 GB | desktop / CLI (default) |
| `knaif-qwen3-1.7b-v1-q6_k.gguf`   | Qwen3-1.7B | 1.7B | Q6_K   | 1.4 GB | mobile / low-footprint |

Public release names are `v1`; the underlying fine-tune was training cycle `sft-v3-flat`.

## Intended use

Structured **plan generation for knaif skills**, not open-ended chat. Given the knaif prompt
(available tools + the user request), the model emits only a `{"plan": [...]}` envelope naming
declared tools and arguments. Unknown tools and unsupported arguments are rejected downstream,
so the model's job is intent → validated plan, not command execution.

## Usage

Via the knaif runtime (recommended — it supplies the prompt and validates/executes the plan):

```bash
knaif models pull qwen3-4b-v1        # or qwen3-1.7b-v1
knaif run ffmpeg "compress holiday.mp4 for whatsapp"