GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
miniweb
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
miniweb
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Task success (%) of Qwen3.5-4B (the paper's Table 3). MiniWeb: the 63 human tasks of the 13 held-out sites, from which no training data comes; the other three benchmarks are not trained on either.
| MiniWeb held-out | WebArena-Lite | WebVoyager | Online-Mind2Web | |
|---|---|---|---|---|
| Untrained single agent | 37.0 | 27.1 | 45.5 | 17.9 |
| Untrained planner + untrained executor | 42.3 | 30.7 | 59.9 | 21.8 |
| MiniWebAgent (untrained planner + this executor) | 54.5 | 36.4 | 65.5 | 33.9 |
On the 103 generated single-primitive tasks of the held-out sites, the trained model as a single agent goes from 55.3% to 80.3%, with gains in every skill family. On whole tasks it needs the planner: trained mostly on one-primitive episodes, the single agent often stops before reaching the skills it learned (37.0% on the held-out tasks, as untrained).
Serve the base model with the adapter (vLLM), then run MiniWebAgent with the repository's evaluator:
bash
python scripts/fetch_adapters.py # -> data/webmix/adapters/pooled_v4MODEL=Qwen/Qwen3.5-4B SERVED=qwen35-4b PORT=8400 MAX_LORA_RANK=64 \ LORA_MODULES=pooled_v4=data/webmix/adapters/pooled_v4 bash src/miniwebagent/scripts/serve.shpython -m miniwebagent.evaluate --human --arm agent_planner --agent-version v6 --planner-name qwen35-4b \ --adapters none --fallback pooled_v4 --vllm http://127.0.0.1:8400/v1 --ports 8300,8301 --out runs/miniwebagentWith PEFT directly: PeftModel.from_pretrained(base_model, "miniweb/MiniWebAgent").
The executor expects the browser-use agent prompt it was trained on (the
system prompt, the task, the agent's memory, the indexed interactive elements and the screenshot) and answers with
its memory and one action. Thinking is off: the chat template renders an empty <think></think> block.
training_steps table, rows with is_target and not
validation): 3,128 verified single-primitive trajectories on the 52 training sites, 857 chains of two trajectories
and 629 trajectories with an injected and corrected mistake, replayed through the agent harness with the base model
writing each step's memory.training_meta.json and train_log.jsonl hold the settings and the loss curve.The paper's other trained executors, in ablations/ (same base model and LoRA shape; each folder has its own
training_meta.json). python scripts/fetch_adapters.py --all downloads them under the names the paper's run commands
use (scripts/reproduce.py show <arm>).
| Folder | What it is | Held-out task success |
|---|---|---|
| (root) | MiniWebAgent's executor | 54.5 |
ablations/task_centric | trained at matched size on task-centric trajectories: InSTA's pipeline reimplemented on the 52 training sites (the dataset's task_centric_steps) | 45.5 |
ablations/specialists/{drag,io,select,rest,reasoning} | one specialist per skill family group (supplementary material): drag, io, select and rest start from the main executor and train on their families' steps with a KL penalty toward it; starts from the base model |
All numbers are with the untrained planner except no_done (single agent; 37.0 with the done steps). The specialists
are served under their family names and picked by the planner's family for each subtask:
LORA_MODULES='drag=.../drag io=.../io select=.../select rest=.../rest reasoning=.../reasoning pooled_v4=...' and
--adapters drag,io,select,rest,reasoning --fallback pooled_v4.
Trained only on MiniWeb's synthetic sites and English instructions, mostly on single primitives: on whole tasks it relies on the planner for their structure. Outside MiniWeb it has been evaluated only on the three benchmarks above.
Apache 2.0, as the base model.
reasoning| 51.3 |
ablations/no_done | the main recipe without each episode's final done step, a control for stopping early (single agent) | 33.9 |