Qwen 3.6 35B Vision — 16 GB VRAM, IQ3_S

A compact multimodal build of Qwen 3.6 35B with vision support, prepared for local use with Ollama.

This model is designed to fit within 16 GB of VRAM, including a very large context window of up to 262,144 tokens, when used with:

• IQ3_S model quantization
• Q4_0 KV-cache quantization
• Flash Attention

Model

byczech/qwen3.6-mmproj-16G-UD-IQ3_S:35b

Highlights

• Text and image understanding
• Qwen 3.6 35B base model
• Multimodal projector included
• IQ3_S quantization
• Up to 262,144 tokens of context
• Optimized for GPUs with 16 GB VRAM
• Suitable for long documents, screenshots, image analysis, coding and general chat

Run

ollama run byczech/qwen3.6-mmproj-16G-UD-IQ3_S:35b

Recommended Ollama configuration

To use the full 262,144-token context while minimizing VRAM usage, start the Ollama server with:

OLLAMA_FLASH_ATTENTION=1 \
OLLAMA_KV_CACHE_TYPE=q4_0 \
OLLAMA_CONTEXT_LENGTH=262144 \
ollama serve

Then run the model in another terminal:

ollama run byczech/qwen3.6-mmproj-16G-UD-IQ3_S:35b

You can verify whether the model is fully loaded on the GPU with:

ollama ps

The PROCESSOR column should report 100% GPU.

Vision example with Python

Install the Ollama Python package:

pip install ollama

Then pass an image path to the model:

from ollama import chat

response = chat(
    model="byczech/qwen3.6-mmproj-16G-UD-IQ3_S:35b",
    messages=[
        {
            "role": "user",
            "content": "Describe this image in detail.",
            "images": ["image.jpg"],
        }
    ],
)

print(response.message.content)

Context length

Ollama may select a much smaller default context on systems with less than 24 GB of VRAM. Set OLLAMA_CONTEXT_LENGTH=262144 (may vary from GPU to GPU try 128K to 256K) before starting the server, or configure num_ctx explicitly in your client (or you can create your own Modelfile with num_ctx parameter).

The maximum context consumes substantially more memory than short prompts. The 16 GB target assumes Q4_0 KV cache, Flash Attention, one loaded model and no parallel requests.

Quality and memory trade-off

Q4_0 KV-cache quantization uses roughly one quarter of the memory required by the default F16 KV cache. This makes the very large context practical on a 16 GB GPU, but it can reduce precision, especially near the maximum context length.

For higher KV-cache precision, use q8_0, but expect higher VRAM usage and a smaller practical context window.

Notes

• Actual VRAM use depends on the Ollama version, GPU backend, prompt length and runtime overhead.
• Close other GPU-intensive applications when using the maximum context.
• Keep OLLAMA_NUM_PARALLEL=1 for the lowest memory usage.
• Image inputs also consume context tokens.
• Newer Ollama versions (0.31+) may need lower context window to fit to 16 GB VRAM - check with ollama ps command