Woof Selector 4B 4-bit v2.1

Choose the next action. Keep browser tasks moving.

Woof Selector is a compact, local browser-use model. Given a task, the current page context and available choices, it selects the next operation and target. Quantization-aware training adapts the model for 4-bit deployment. Woof Selector is built on Qwen3.5-4B through browser-focused post-training.

The selector decides what to do and where; the text filler supplies what to type.

Single-step validation

The QAT model achieved 96.6% decision acceptance (516/534) on fixed browser-state validation, evaluated using an LLM-as-judge. Valid alternative actions were accepted.

End-to-end browser evaluation

10 public tasks

Woof runs use our selector and text fillers on Apple Silicon. Original Qwen results are also our measurements. Other models' results are reported by AIMultiple on the same 10 public tasks.

Model	Successful Tasks	Success Rate	Median Time per Task
Astra	9/10	90%	18.30 s
Gemini 3.8 Flash	9/10	90%	12.60 s
Woof-4B selector + Woof-4B textfiller	7/10	70%	10.68 s
Woof-4B selector + Woof-0.8B textfiller	7/10	70%	10.94 s
Jev 1.13	5/10	50%	6.45 s
Kev-9B	5/10	50%	43.40 s
Solar Decide	5/10	50%	22.90 s
Qwen3.5-4B selector + Qwen3.5-4B textfiller	4/10	40%	16.29 s
Qwen3.5-4B selector + Qwen3.5-0.8B textfiller	4/10	40%	15.01 s
Kev-4B	3/10	30%	7.15 s
Kev-4B via OpenRouter	2/10	20%	3.80 s
Bespoke Nimble 9B	1/10	10%	2.55 s
Tev1-4B	1/10	10%	4.80 s
Tev1-0.8B	0/10	0%	1.80 s
CLM-v0.1-8B	0/10	0%	8.45 s
Laya	0/10	0%	21.35 s
GLiNER2.5-Decide	0/10	0%	0.35 s

Both Woof combinations achieved 70% task success with a locally hosted 4B browser-use model, including the combination with the lightweight 0.8B text filler.

54 live-web tasks

We evaluated DOM-only browser tasks covering flights, accommodation, shopping and restaurant searches, with task completion assessed using an LLM-as-judge.

Model	Successful Tasks	Success Rate	Success Rate Excluding Anti-bot Skips
Woof-4B	26/54	48.1%	51.0%
Kev-9B Browser Post-trained	18/54	33.3%	38.3%
Qwen3.5-4B	9/54	16.7%	18.4%
Kev-9B	8/54	14.8%	16.0%

Download the task definitions and evaluation criteria. Evaluation methods are described in Evaluation details.

Usage

Download mlx/ for Apple Silicon or gguf/ for a compatible GGUF runtime. Download only the variant you need and keep its files together.

Use a compatible Woof Selector harness in Underdog to provide the task, page observations and candidate choices. The harness executes the selected action, reads the updated page and asks for the next decision. The model runs locally after download.

Pair Woof Selector with one of our text fillers when an action requires generated text:

Text Filler	Apple Silicon / MLX 4-bit	GGUF 4-bit
Woof-4B textfiller	Download	Download
Woof-0.8B textfiller	Download	Download

Both variants use 4-bit weights. MLX uses BF16 computation; GGUF uses Q4_K_M weights with runtime-selected computation. Enable Flash Attention for the GGUF selector. Preserve the supplied configuration, tokenizer and chat template.