The vision encoder is a second set of weights that sits in GPU memory next to the language model. For Qwen3-VL-8B at Q4_K_M it adds 1.78 GiB, taking an 8K-context run from 5.93 to 6.48 GiB to 7.71 to 8.26 GiB. That small jump is enough to push it off an 8 GB card. For Qwen2.5-VL-7B the Unsloth vision file is 5.04 GiB, larger than the text model itself.
The vision part does not change the KV cache. It is a fixed block of weights, loaded once, whether you send one image or none. (Image tokens do take up context like any other tokens.)
Vision off vs vision on
| Model (Q4_K_M) | Text weights (bytes) | Vision component (bytes) | Text only (GiB) | With vision (GiB) | 8 GB RTX 4060 with vision | 12 GB RTX 4070 with vision |
|---|---|---|---|---|---|---|
| Qwen3-VL-8B | 5,027,784,800 | 1,911,319,552 | 5.93 to 6.48 | 7.71 to 8.26 | won’t fit | fits |
| Qwen2.5-VL-7B | 4,683,072,384 | 5,411,547,680 | 4.92 to 5.42 | 9.96 to 10.46 | won’t fit | tight |
8K context (8,192 tokens), one chat, FP16 KV cache, llama.cpp. Usable memory is 7 GiB on the RTX 4060 and 11 GiB on the RTX 4070 after the default 1 GiB reserve. Qwen3-VL-8B is the official Qwen/Qwen3-VL-8B-Instruct-GGUF build (Qwen3VL-8B-Instruct-Q4_K_M.gguf); Qwen2.5-VL-7B is unsloth/Qwen2.5-VL-7B-Instruct-GGUF.
Both fit easily on a 24 GB RTX 4090 with vision on. The text KV cache for Qwen3-VL-8B is 1.13 GiB at 8K in either mode.
Practical takeaway: if you have an 8 GB card and want images, Qwen3-VL-8B at Q4_K_M won’t load with an 8K context. Text-only, it is tight. On 12 GB it fits with vision. See Qwen3-VL-8B on an RTX 4070 for the breakdown.
Different builds, different vision sizes
The Qwen3-VL-8B Q5_K_M and Q6_K files in the catalog are Unsloth builds, and their vision component is a different size (4,627,174,080 bytes). The calculator doesn’t swap one publisher’s vision file into another’s quant; the table uses the official Q4_K_M artifact. Check the vision file size of the exact build you download.
The BF16 checkpoint
The official Qwen/Qwen3-VL-8B-Instruct BF16 shards add up to 17,534,247,392 bytes. The vision encoder lives inside those shards instead of in its own file, so its separate size is unknown. With vision on, the calculator marks the estimate as partial and returns 17.58 to 21.11 GiB at 8K context, padding the upper bound to cover the missing split. Without that pad the whole-model range is 17.58 to 18.36 GiB.
Numbers are estimates from model configs and published file sizes, not measurements. How we calculate.
Related: Gemma 3 ships a smaller vision projector and uses a sliding-window cache, covered in Gemma 3 sliding-window VRAM. The text-only Qwen3-8B (a different checkpoint) is in Qwen3 VRAM by size. Model pages: Qwen3-VL-8B-Instruct, Qwen2.5-VL-7B-Instruct.
FAQ
How much extra VRAM does vision add to Qwen3-VL-8B?
1.78 GiB for the official Q4_K_M build (a 1,911,319,552-byte vision component). At 8K context the total goes from 5.93 to 6.48 GiB text-only to 7.71 to 8.26 GiB with vision.
Can I run Qwen3-VL-8B with images on an 8 GB GPU?
Not at Q4_K_M with an 8K context: with vision it needs 7.71 to 8.26 GiB, over the 7 GiB usable on an RTX 4060. Text-only is tight but loads. A 12 GB card such as the RTX 4070 fits it with vision.
Can Qwen2.5-VL-7B Q4_K_M fit an 8 GB card with vision?
No. Its Unsloth vision component is 5.04 GiB, bigger than the text weights, so the total is 9.96 to 10.46 GiB. Text-only it fits the same card at 4.92 to 5.42 GiB.
Why is there no vision size for the BF16 Qwen3-VL-8B checkpoint?
The official safetensors shards don't keep the vision encoder in a separate file, so there is no published byte count to use. The calculator marks that estimate as partial and pads the upper bound instead of guessing a split.
Numbers from calculator 0.3.0, model data 2026-10-02.1. Sources last checked 2026-10-29.