knaif — natural language → validated action plans

Fine-tuned Qwen3 models for knaif: 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):

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