5 models · Vision + language · Transformers.js
Vision-language models in the browser
5 models handle vision + language in the browser today, from 180.1 MB (SmolVLM-256M-Instruct) to 616.9 MB (Qwen3.5-0.8B), a 252.1 MB download at the midpoint. Every one runs with Transformers.js over WebGPU or WebAssembly: no server, no install, no account.
- Models
- 5
- Smallest
- 180.1 MB
- Median
- 252.1 MB
- Largest
- 616.9 MB
Every vision + language model, ranked by WebGPU size
| Model | WebGPU download | WASM download |
|---|---|---|
| SmolVLM-256M-Instruct 256.5M params | 180.1 MB q4f16 | 247.8 MB uint8 |
| Florence-2-base-ft 231.6M params | 213.1 MB q4f16 | 262.2 MB uint8 |
| Granite-Docling-258M 257.52M params | 252.1 MB q4f16 | 450.2 MB q8 |
| Florence-2-large-ft 770.43M params | 555.3 MB q4f16 | 791.1 MB uint8 |
| Qwen3.5-0.8B 873.44M params | 616.9 MB q4f16 | 1.20 GB q8 |
Sizes measured from the HuggingFace API file tree for each model's repo, not estimated. Sorted smallest to largest by the WebGPU headline pick.
Size ladder
The WebGPU download spans 3.4x here: 180.1 MB (SmolVLM-256M-Instruct) to 616.9 MB (Qwen3.5-0.8B).
None of the 5 models here download smaller over WebAssembly than WebGPU: the WebGPU build is the lighter pick across the group.
None of the 5 models here fit under 100 MB over WebGPU.
A bytes-per-parameter comparison needs at least 3 models with a verified parameter count and a real spread between them; this group does not clear that bar yet.
Notable vision + language models
Other tasks
Or see every browser model grouped by task, or the main GGUF catalog for native runtimes outside the browser.
FAQ
Do all vision + language models here use the same WebGPU quant?
Yes: all 5 models here use the q4f16 WebGPU build.
What would it cost to download every vision + language model here?
1.77 GB total over WebGPU across all 5 models, if you tried every one back to back. Most projects only need the single model that fits the job, not the whole set.
Sizes measured 2026-08-01 from the HuggingFace API. Last validated 2026-09-14. See all browser models or the methodology.