Qwen3.5

Use Alibaba's Qwen3.5-VL vision-language model through Workflows, Dedicated Deployments, or self-hosted Inference

Qwen3.5 is Alibaba's vision-language model family. It accepts an image and a text prompt and returns a text response. Two pretrained checkpoints are available:

Alias	Parameters
qwen3_5-0.8b	0.8B
qwen3_5-2b	2B

Qwen3.5 accuracy

Headline vision-language benchmarks (non-thinking mode) from the official model cards (0.8B, 2B):

Benchmark	qwen3_5-0.8b	qwen3_5-2b
MMMU	47.4	64.2
MathVista (mini)	58.6	73.9
MMBench (EN v1.1)	68.0	81.3

Qwen3.5 inference speed

Latency measured with Roboflow Inference on 1x NVIDIA L4, batch size 1, generating exactly 128 tokens with greedy decoding from a fixed prompt. Latency scales with output length, so use tokens/sec to estimate other lengths.

Alias	Latency, 128 tokens (ms)	Tokens/sec
qwen3_5-0.8b	3307	39
qwen3_5-2b	3688	35

Use Qwen3.5 in a Workflow

Qwen 3.5 VL is available as a preconfigured Workflow on the "Open-Source Models" tab of the Models page. Select "Qwen VL", choose a model variant and prompt, then click "Test API" to fork the Workflow into your Workspace and start running inference.

The Workflow uses the unified qwen_vlm@v1 block, which supports multiple Qwen VL generations:

Model	Parameters
Qwen 3.5 VL 0.8B	0.8B
Qwen 3.5 VL 2B	2B
Qwen 3 VL 2B	2B
Qwen 2.5 VL 7B	7B

Qwen3.5 API

Direct Inference SDK calls to Qwen3.5 require a Dedicated Deployment or self-hosted Inference. For hosted access, use the Workflow path described above.

1

Get your API Key

Create a Roboflow account, find your key on the Roboflow API settings page and make it available to your shell:

export ROBOFLOW_API_KEY="your-key-here"

2

Install the dependencies

Install the Inference SDK:

pip install -U inference-sdk supervision

3

Run the model

Set api_url to your Dedicated Deployment URL or a local Inference server.

import os
import supervision as sv
from inference_sdk import InferenceHTTPClient, InferenceConfiguration

image = sv.load_image_from_url("https://media.roboflow.com/quickstart/dog.jpeg")

client = InferenceHTTPClient(
    api_url="https://your-deployment.roboflow.cloud",
    api_key=os.environ["ROBOFLOW_API_KEY"],
).configure(InferenceConfiguration(api_key_transport="header"))
result = client.infer_lmm(
    image,
    model_id="qwen3_5-2b",
    prompt="Describe this image briefly.",
    max_new_tokens=256,
)
print(result["response"])

The code above prints the model response to the terminal:

A person wearing a white t-shirt and red shorts is carrying a black backpack on their shoulder, with a beagle dog perched on top of it. The scene takes place outdoors in a residential area, with modern apartment buildings in the background and greenery along the sidewalk. The person appears to be walking or standing near a building with large windows.

Set api_url to match your deployment target:

• http://localhost:9001 for a local Inference server.

• Your Dedicated Deployment URL for a private endpoint.

You can train your own Qwen3.5 checkpoint on Roboflow and call it by its per-model {workspace}/{model-slug} ID (see Versions, Trainings, and Models).

Run Qwen3.5 with self-hosted Inference

Qwen3.5 can also be loaded directly with the inference package instead of being called over HTTP.

1

Install the package

pip install "inference[transformers]"

Use inference-gpu[transformers] on a GPU machine.

2

Run the model

from inference.models.qwen3_5vl.qwen3_5vl_inference_models import (
    InferenceModelsQwen35VLAdapter,
)

model = InferenceModelsQwen35VLAdapter(
    model_id="qwen3_5-0.8b",
    api_key="YOUR_API_KEY",
)

image = "https://media.roboflow.com/dog.jpeg"
prompt = "How many dogs are in this image?"

preprocessed, metadata = model.preprocess(image, prompt)
predictions = model.predict(preprocessed)
result = model.postprocess(predictions, metadata)

print(result[0].response)

Qwen3.5 also supports a "thinking" mode, in which the model generates reasoning tokens before answering.

Execution modes in Workflows

When used in a Workflow, Qwen3.5 runs in one of two modes:

• Local execution: the model runs on your Inference server (GPU recommended).

• Remote execution: the model is invoked over HTTP on a remote Inference server.