gloofy-1-nano

An open language model for marketing work, built on Qwen3-4B (Apache 2.0) and fine-tuned by overads Inc.

It tags ad creative against a closed taxonomy, qualifies leads, and writes video hooks, where it beat the model it was built from 36 to 9 under blind pairwise judging. It runs offline on a laptop, so the marginal cost of the ten-millionth call is the same as the first, and no ad copy leaves your machine.

What it does NOT ship for. Campaign diagnosis, campaign setup, caption length discipline and image direction were each measured against the untrained Qwen3-4B base, and the base won all four. Those jobs ship as the plain base model with a task system prompt, no adapter. The rule we take from it: fine-tuning wins where the knowledge is tacit and compression-shaped, and loses where the knowledge can simply be stated in the prompt, because training on stateable facts costs instruction-following. Full per-family numbers are in the benchmark.

Scores, including where it loses

Measured on 150 REAL published ads, every model run through one harness with one specification and one scorer.

model	exact facet accuracy
Claude	0.916
annotator agreement ceiling	0.911
Gemini 3.6 Flash	0.817
GPT-5.5	0.792
MLX adapter (this model, unquantised)	0.780
this GGUF, Q4_K_M	0.735
Qwen3-4B untrained	0.000

Read the GGUF row, not the adapter row. The file in this repo scores 0.735. The 0.780 is the MLX adapter before export, and quantising to Q4_K_M costs 0.045, concentrated in the judgment facets (angle -0.100, hook -0.060) while extraction facets barely move. Publishing the higher number for a file that earns the lower one would be dishonest, so both are here.

And read them with a margin. Three training runs differing only by random seed spread 0.055 on this exam. Frontier models genuinely beat this one; the gaps between adjacent rows below Claude are less certain than they look.

Reproduce every number

The benchmark is public and was released before this model, so the questions could not be tuned to suit the answers: github.com/overads-Inc/gloofy_bench

llama-server -m gloofy-1-nano-Q4_K_M.gguf --port 8080

uv run python harness/run_frontier.py --provider local \
  --base-url http://127.0.0.1:8080/v1/chat/completions \
  --model gloofy-1-nano --exam tasks/exam-v3.1.jsonl --out r.jsonl
uv run python harness/score_dump.py r.jsonl

Usage

llama-cli -m gloofy-1-nano-Q4_K_M.gguf -p "<your prompt>"

Or with ollama, using the Modelfile in this repo:

ollama create gloofy-1-nano -f Modelfile
ollama run gloofy-1-nano

The model expects the taxonomy in its system prompt. The exact specification used for every score above is in harness/run_frontier.py.

What it is not

• Not a frontier model. Claude, Gemini and GPT-5.5 all beat it on our own exam and the table above says so.
• Not good at open reasoning. On campaign diagnosis it lost 9-1 in blind judging to the untrained base model, so that job should be routed to a general model instead.
• Not a replacement for judgment. Annotators agree with each other only 91% of the time on these labels, which caps what any model trained on them can achieve.

How it was built

21 published eval cards document the whole thing, including the failures: a plateau across six rounds, four conclusions retracted when a variance test proved them noise, a "human ceiling" that turned out to be three Claude agents, an exam that never listed its own field names, and 9.8 points found hiding in prompt formatting.

Attribution

Built on Qwen3-4B by Alibaba, Apache 2.0. Fine-tuned by overads Inc. Rebranding is permitted by that licence; concealing the base would contradict everything else published here.