Instructions to use HaoranLiu/SFT-4B-ScaleCUA-MixedOpen with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use HaoranLiu/SFT-4B-ScaleCUA-MixedOpen with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="HaoranLiu/SFT-4B-ScaleCUA-MixedOpen")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("HaoranLiu/SFT-4B-ScaleCUA-MixedOpen")
model = AutoModelForMultimodalLM.from_pretrained("HaoranLiu/SFT-4B-ScaleCUA-MixedOpen", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use HaoranLiu/SFT-4B-ScaleCUA-MixedOpen with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "HaoranLiu/SFT-4B-ScaleCUA-MixedOpen"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HaoranLiu/SFT-4B-ScaleCUA-MixedOpen",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/HaoranLiu/SFT-4B-ScaleCUA-MixedOpen

SGLang

How to use HaoranLiu/SFT-4B-ScaleCUA-MixedOpen with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "HaoranLiu/SFT-4B-ScaleCUA-MixedOpen" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HaoranLiu/SFT-4B-ScaleCUA-MixedOpen",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "HaoranLiu/SFT-4B-ScaleCUA-MixedOpen" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HaoranLiu/SFT-4B-ScaleCUA-MixedOpen",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use HaoranLiu/SFT-4B-ScaleCUA-MixedOpen with Docker Model Runner:
```
docker model run hf.co/HaoranLiu/SFT-4B-ScaleCUA-MixedOpen
```

SFT-4B-ScaleCUA-MixedOpen

Qwen3-VL-4B-Instruct supervised-finetuned on 1430 Lite.ScaleCUA trajectories from three open-source teachers in equal proportion, trained with cua-lite + slime.

This is the epoch-3 checkpoint (iter_1070) — the best of the ScaleCUA SFT arms measured here.

Results — Lite.OSWorld `eval` split

332 tasks, greedy (temperature=0), max_steps: 30, concurrency 12, group_size=1, mean over all 332 with invalid tasks scored 0. Every row below reached Valid 332/332 on one host under one protocol, so the rows are comparable to each other.

All rows are epoch-3 (final) checkpoints.

model	teachers	trajs	mean episode_return	success
`SFT-4B-ScaleCUA-MixedOpen` (this)	3 open-source, 1/task	1430	0.3927	125/332
`-MixedOpen-cap2`	same 3, up to 2/task	2520	0.3802	120/332
`-Qwen35-27B`	Qwen3.5-27B only	—	0.3682	117/332
`-Qwen38`	Qwen3.8-27B only	1224	0.3623	115/332
gpt-5.5 arm	gpt-5.5 only	1254	0.3241	103/332
`-EvoCUA`	EvoCUA-8B only	865	0.3215	101/332

The gpt-5.5 row uses the unmodified gpt-5.5 cohort. An earlier gpt-5.5 collection reached 0.3753 with prompt engineering applied to the teacher, which is not comparable to the other rows here — every other teacher was collected without it, so that number is excluded rather than listed.

Three things the table settles:

Mixing beats every single teacher. Against the strongest single-teacher arm (Qwen3.5-27B) the mix is +0.0245 / +8 tasks, and against Qwen38 — the closest match in pipeline and trajectory count — +0.0304 / +10. Both clear the threshold used here for a real effect (>0.02 mean and ≥7 tasks). Caveat below: teacher identity and task coverage move together, so this is not a clean single-variable ablation.

A second trajectory per task does not help. cap2 doubles the data (2520 vs 1430) at matched epochs and lands −0.0125 / −5 tasks below this model. Ten of its per-domain cells move, five down and two up; the direction is consistently mildly negative. More solutions for a task the model already has an example of is not where the headroom is.

Two of the three teachers carry the mix. Qwen3.5-27B and Qwen38 land within 0.006 of each other; EvoCUA-8B alone is 0.0408 / 14 tasks below Qwen38 on only 865 trajectories. A two-teacher mix that drops EvoCUA is the obvious next arm.

Per-domain (this checkpoint)

domain	n	mean	pass
`vs_code`	18	0.722	13
`thunderbird`	14	0.643	9
`libreoffice_writer`	22	0.591	12
`vlc`	15	0.587	7
`os`	19	0.526	10
`gimp`	16	0.500	8
`chrome`	43	0.487	20
`libreoffice_impress`	47	0.448	21
`libreoffice_calc`	46	0.261	12
`multi_apps`	92	0.159	13

multi_apps is 92 of 332 tasks (28%) and by far the weakest — the same shape the single-teacher arms show. Multi-application coordination is the open problem in this task pool, and no teacher mix measured here moves it.

Data

Equal-proportion mix over lite.scalecua rl tasks, one trajectory per task:

teacher	trajectories	share
`Qwen/Qwen3.8-27B`	478	33.4%
`Qwen/Qwen3.5-27B`	477	33.4%
`meituan/EvoCUA-8B-20260105`	475	33.2%

1430 unique tasks (one trajectory each, no task appears twice). 1390/1430 are full successes (episode_return == 1.0); the rest score 0.6–0.67 partial credit. All rows come from the rl_tasks split.

The premise of the mix is that different agents specialize in different domains, so pooling open-source teachers widens task coverage. Note this means the mixed arm sees more tasks, not merely more trajectories — coverage and teacher diversity move together here and are not separated by this experiment.

Training

base            Qwen/Qwen3-VL-4B-Instruct
export config   scripts/configs/qwen3_vl/default/lite.osworld.yaml  (full_history_size=4)
objective       token-level SFT on assistant-action tokens; screenshots are
                context only (masked out of the loss)
lr              5e-6 cosine -> min_lr 1e-6, warmup_fraction 0.1
batch           global_batch_size 4 trajectories/step, micro_batch_size 1
epochs          3  (1070 steps, save_interval 358)
parallelism     TP=2, 2x 80GB H100 -> DP=1
trained         2026-09-10 05:04 -> 20:52 UTC (15h47m)
code            cua-lite-preference @ 6fbb2c9b, slime v0.3.0
ray_job         raysubmit_TxuEdEMpVAAwC22Y
wandb           liu-hr22-tsinghua-university/cua-lite-dev/runs/8mpvuigx

Sibling checkpoints kept locally: iter_356/iter_357 (ep1), iter_713/iter_715 (ep2).

Usage

uv run python scripts/rollout.py \
  --model-id Qwen/Qwen3-VL-4B-Instruct \
  --model-path <path to this checkpoint> \
  --env-id lite.osworld --splits eval \
  --filter "lambda m: not m.others.get('exclude_reason')" \
  --concurrency 12 \
  --config-path scripts/configs/qwen3_vl/default/lite.osworld.yaml

The qwen3_vl adapter/config is required: the model was trained on that history protocol's rendering (full_history_size=4). Serving it under a different window silently changes the prompt the model was fit to.

Caveats

Only compare numbers produced on one host in one protocol. The same checkpoint has shifted by up to 0.018 / 7 tasks between VMs here — enough to flip a "real effect" verdict. The table above is one host, one protocol.
Not a single-variable ablation. Mixed vs single teacher differs in teacher identity, trajectory count, and task coverage simultaneously.
num_valid first. summary.json's mean_episode_return averages over valid tasks only, and a run may write an interim summary before its retry round finishes. Check Valid N/332 before reading any mean.
Trained on rl tasks of lite.scalecua; evaluated on lite.osworld eval. These are separate task pools on a shared desktop substrate.

Downloads last month: 29

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for HaoranLiu/SFT-4B-ScaleCUA-MixedOpen

Base model

Qwen/Qwen3-VL-4B-Instruct

Finetuned

(446)

this model

HaoranLiu
/

SFT-4B-ScaleCUA-MixedOpen

SFT-4B-ScaleCUA-MixedOpen

Results — Lite.OSWorld `eval` split

Per-domain (this checkpoint)

Data

Training

Usage

Caveats

Model tree for HaoranLiu/SFT-4B-ScaleCUA-MixedOpen

Dataset used to train HaoranLiu/SFT-4B-ScaleCUA-MixedOpen

SFT-4B-ScaleCUA-MixedOpen

Results — Lite.OSWorld eval split

Per-domain (this checkpoint)

Data

Training

Usage

Caveats

Model tree for HaoranLiu/SFT-4B-ScaleCUA-MixedOpen

Dataset used to train HaoranLiu/SFT-4B-ScaleCUA-MixedOpen

Results — Lite.OSWorld `eval` split