--- title: "Qwen 3.6 - 35B Vision Language model" url: https://docs.catalyst.zoho.com/en/quickml/help/available-models/qwen-3.6-35b-vision-language/ description: "QuickML is a fully no-code ML pipeline builder service in the Catalyst development platform for creating machine-learning pipelines with end-to-end solutions." last_updated: 2026-08-18T06:07:37Z service: QuickML type: general --- **VL-Qwen3.6-35B-A3B** is a 35-billion-parameter Multimodal Mixture-of-Experts (MoE) model with approximately 3 billion active parameters, deployed in 8-bit precision. It is designed for combined text and image understanding, and offers advanced reasoning, Agentic coding, and thinking-preservation capabilities. The notation **35B-A3B** means the model has 35 billion total parameters, but only ~3 billion are activated per token during inference. It therefore computes like a 3B model per token, keeping latency and cost low despite its large total capacity. Optimized for Multimodal understanding, Advanced reasoning capabilities, Thinking preservation and Agentic coding workflows. It delivers performance comparable to larger dense models at a fraction of the compute cost, making it practical for cost-sensitive and small-scale deployments. ### When should you use Qwen 3.6 - 35B Vision Language? Qwen 3.6 - 35B Vision Language is best suited for: - Multimodal tasks combining text and images, such as document understanding, chart and diagram interpretation, and image-based question answering. - Small-scale or cost-efficient inference where a full 35B dense model would be impractical. - Cost-sensitive deployments requiring low per-token compute. ### Supported input and output formats **Input type**: Text and Image **Output formats**: Text **Model Key**: `VL-Qwen3.6-35B-A3B` > **Note:** Audio/Video input is not supported. The model accepts text and image inputs only. ### Capabilities - Natively Multimodal — meaning it is multimodal by design, instead of just an add-on layer. - Advanced perception capability that far exceeds the relative performance of other models of similar size with only 3 billion activated parameters. - Multimodal reasoning across modalities like documents, images, and charts with the instruction prompt combined. - Efficient multi-step reasoning ability across complex reasoning tasks. - Retains thinking context across conversation history. ### Model details To view the model details, go to LLM Serving, select the **Models** tab, and choose the Qwen 3.6 - 35B Vision Language model. - **Model Size**: 35 billion (35B) total parameters, providing large capacity for multimodal understanding and generation. - **Active Parameters**: Although the model holds 35 billion parameters in total, it only activates about 3 billion of them to process each word (token). This keeps inference fast and efficient without sacrificing quality. - **Precision**: Deployed in 8-bit precision to reduce memory footprint and serving cost. - **Architecture**: The Mixture-of-Experts design routes each token to a sparse subset of expert layers, enabling large total capacity while keeping per-token compute low. - **Input Context Window**: Supports inputs up to three images (6k tokens total) plus text (3k tokens), for approximately 9k tokens overall. In a Mixture-of-Experts model, the parameters are split into groups called "experts." For each token, only a small number of these expert groups are selected and used. This reduces the compute per token while keeping the total model capacity large. ### Model parameters QuickML provides a robust and flexible LLM (Large Language Model) serving environment, enabling developers and businesses to fine-tune AI model behavior through a variety of customizable parameters. Unlike other platforms that restrict user control, QuickML prioritizes transparency and adaptability—allowing you to balance creativity with precision, determine the ideal response length, and guide the model to behave in ways that best serve your domain-specific needs. The Qwen 3.6 - 35B Vision Language model exposes a set of advanced generation parameters in the LLM Serving playground. These parameters give you fine-grained control over the model's instructions and output behavior — including randomness, token selection, and response length. Whether you're extracting insights from documents, using a visual content analyzer, or answering Q&A about images, QuickML's parameter configuration options ensure your application delivers responses that are not only relevant, but also contextually tailored to your multi-modal inputs. Here's a breakdown of the available parameters:
Parameter Description Values / Notes
Model The VLM used for serving Qwen 3.6 35B A3B VLM
Instructions System-level instruction that guides the model's behavior, tone, and response style across the session. This enhances output relevance and consistency for specific application needs. For example, a legal firm can enter instructions like "Provide responses in a formal legal tone with citations where applicable.", ensuring AI-generated content aligns with compliance and professional standards. Free text (Optional)
Temperature Controls randomness in the output. Lower values produce focused, deterministic responses; higher values produce more varied, creative responses. For example, a financial institution setting Temperature to 0.2 would get precise responses like "The Federal Reserve increased interest rates by 0.25%." while setting it to 0.8 might generate "The Federal Reserve's recent rate hike of 0.25% aims to curb inflation, affecting mortgage and loan rates." 0.0 (Focused) – 1.0 (Creative); default 0.7
Top-K Limits next-token selection to the K most probable words, controlling how much vocabulary the model considers at each step. For example, a corporate HR chatbot with Top-K set to 10 may generate standard responses like "We value diversity in our hiring process.", whereas 50 allows for richer, more engaging responses like "At [Company Name], diversity is at the core of our hiring process, fostering innovation and inclusivity." 1 – 100; default 50
Top-P Cumulative probability threshold for selecting the next token (nucleus sampling); the model considers the smallest set of tokens whose combined probability meets this value. For example, a customer support AI using a Top-P setting of 0.3 might generate straightforward responses like "Your order will arrive in 3 days.", while 0.9 could lead to "Your order is expected to arrive within 3 days. You'll receive a tracking update soon! Let us know if you need further assistance." 0.0 – 1.0; default 0.9
Max Tokens Defines the maximum number of tokens (between 1 and 4096) that the model can generate in a response. QuickML allows precise control over response length, optimizing cost and latency. Numeric; limit varies by model (up to 4096 for GLM 4.7 Flash); default 500
> **Note:** Default values are suitable for most use cases; adjust them when you need to tune response quality, creativity, or length for a specific application. ### Access the model The **Qwen 3.6 - 35B Vision Language** model can be accessed by creating an LLM Serving . Refer to the below help documentation to understand the process of creating endpoints for LLM : - LLM Serving — [LLM Endpoints](https://docs.catalyst.zoho.com/en/quickml/help/endpoints/llm-serving-endpoint/) Once the endpoint is created, API details and the relevant SDK information will be provided for integration in downstream applications. #### 1. API Details The following details apply to the endpoint created for this model. - **Endpoint URL**: The Model API URL used to send prompts to the model. - **OAuth Scope**: The permission level required to access the model. - **Authentication**: Specifies OAuth as the method for verifying user identity. - **HTTP Method**: Indicates that API requests must be made using the POST method. - **Headers**: Requires metadata, including organization ID and OAuth token, for authorization. - **Sample Request**: A sample input JSON showing how to send a prompt to the Qwen 3.6 - 35B Vision Language model along with parameters such as top_p, temperature, and max_tokens. ```json { "prompt": "Fields to extract: Contact details, Skills, education and project details and give the results in the json format with keys as the above fields to extract and values as the original value in the document image", "images": [ "", "" ] } ``` - **Sample Response**: The output JSON format containing the model's generated response based on the prompt. ```json { "request_id": "a9df919e72da4ef1a1a9360fdb025641", "model": "VL-Qwen3.6-35B-A3B", "response": "```json\n{\n \"Contact\": {\n \"Email\": \"person.one@email.com\",\n \"Phone\": \"+91 98 7654 3210\",\n \"Address\": \"H#1, Area - Chennai 600001\"\n },\n \"Skills\": [\"Windows\", \"HTML5\", \"UI/UX DESIGN\", \"Communication\", \"MS Office\", \"CSS3\", \"WordPress\", \"Teamwork\", \"Internet\", \"JAVASCRIPT\", \"Python\", \"Project Mgmt.\"],\n \"Education\": {\n \"Degree\": \"Bachelor of Engineering in Computer Science\",\n \"Institute\": \"National Institute of Technology, Chennai\",\n \"Year\": \"Aug, 2023\"\n },\n \"Projects\": [\n \"Built a responsive e-commerce website for a local startup, resulting in a 15% increase in online sales.\",\n \"Contributed bug fixes and feature enhancements to the Bootstrap CSS framework.\",\n \"Personal portfolio website showcasing coding skills and past projects.\",\n \"Volunteered for Hack for Education, creating an e-learning platform for underprivileged children.\"\n ]\n}\n```", "metrics": { "input_text_token_length": 57, "input_image_token_length": 1750, "input_guided_prompt_length": 0, "output_text_token_length": 256, "queue_wait_time": 2.4174728393554688, "processing_time": 6.480634450912476, "total_time_taken": 8.898107290267944 } } ``` > **Note:** The model served is fixed when the endpoint is created, so the model field in the request body is optional and does not change which model responds. As this is a vision-language model, image inputs must be passed as additional content items within the message; refer to the API documentation for the exact Multimodal payload format. #### 2. SDK Integration In addition to calling the endpoint directly over REST, you can invoke it programmatically using the Catalyst SDKs, which handles authentication, trigger requests, and response parsing for you. QuickML endpoints are executed through a two-step pattern: create a QuickML component instance, then call the `analyzeImage()` method, passing the endpoint key (the unique ID of the published endpoint, copied from the Catalyst console) and your input data as key-value pairs. The method returns the model's output as a structured response. The SDK is available across languages, refer to the relevant guide for setup, authentication configuration, and method syntax: - **JavaScript** — [Overview](https://docs.catalyst.zoho.com/en/sdk/javascript/v1/quickml/overview/) · [Create QuickML Instance](https://docs.catalyst.zoho.com/en/sdk/javascript/v1/quickml/get-instance/) · [Execute QuickML Endpoint](https://docs.catalyst.zoho.com/en/sdk/javascript/v1/quickml/execute-endpoint/) - **Java** — [Execute QuickML Endpoints](https://docs.catalyst.zoho.com/en/sdk/java/v1/quickml/execute-quickml-endpoints/) - **Python** — [Python SDK documentation](https://docs.catalyst.zoho.com/en/sdk/python/v1/overview/) For JavaScript, install the `@zcatalyst/quickml` package and import the QuickML module before using the SDK methods. Executing an endpoint requires the Admin scope. > **Note:** The model's endpoint must be published in your Catalyst project before the SDK method can be used. QuickML is currently available to Catalyst users in the US, IN, EU, and CA data centers.