---
license: apache-2.0
library_name: streamembed
base_model:
  - Qwen/Qwen3-VL-Embedding-2B
  - Qwen/Qwen3-VL-Embedding-8B
  - google/siglip2-so400m-patch16-naflex
  - google/siglip2-base-patch16-naflex
  - vultr/VultronRetrieverFlash-Qwen3.5-0.8B
tags:
  - screenshots
  - retrieval
  - lora
  - qwen3-vl
  - siglip2
---

# StreamEmbed models

LoRA adapters for [StreamEmbed](https://pypi.org/project/streamembed/), which searches your own screen history by describing what you saw. Only the adapters are stored here; each base model comes from its own repository, pinned to the commit the adapters were trained against.

| Model | Base model (commit) | Size | StreamEmbed support |
|---|---|---|---|
| `qwen3-vl-2b` (default) | [Qwen/Qwen3-VL-Embedding-2B](https://huggingface.co/Qwen/Qwen3-VL-Embedding-2B) (`9f2f7e7`) | 2B | yes |
| `qwen3-vl-8b` | [Qwen/Qwen3-VL-Embedding-8B](https://huggingface.co/Qwen/Qwen3-VL-Embedding-8B) (`2c45655`) | 8B | yes |
| `siglip2-so400m` | [google/siglip2-so400m-patch16-naflex](https://huggingface.co/google/siglip2-so400m-patch16-naflex) (`cc24074`) | 1.1B | weights only |
| `vultron-0.8b` | [vultr/VultronRetrieverFlash-Qwen3.5-0.8B](https://huggingface.co/vultr/VultronRetrieverFlash-Qwen3.5-0.8B) (`6f6b675`) | 0.8B | weights only |
| `siglip2-base` | [google/siglip2-base-patch16-naflex](https://huggingface.co/google/siglip2-base-patch16-naflex) (`b53b807`) | 0.4B | weights only |

StreamEmbed's stream encoder runs Qwen3-VL models. The SigLIP2 and Vultron adapters are published for use with your own code; StreamEmbed lists them but does not run them yet. The SigLIP2 adapters were trained with NaFlex inputs of at most 1024 patches (`max_num_patches` in `models.json`).

Each model folder has two LoRAs (rank 32, alpha 32) on the attention layers of the image and text (or language model) encoders:

- `query_lora.safetensors`: the retrieval LoRA, trained to find screenshots from a description. It encodes questions, and full screenshots when nothing is skipped.
- `frame_adapter.safetensors`: encodes screenshots when repeated blocks are skipped and unchanged ones are reused across frames, so that these vectors stay close to full-screenshot vectors.

`models.json` lists every model with its base repository, commit and files.

## Use

```python
import streamembed as se

se.configure(model="qwen3-vl-2b")
se.build_index_from_directory("screenshots")
results = se.query_index("the flight search page with fares to Denver")
se.visualize_results(results)
```

To apply an adapter yourself, load it into a PEFT model of its base:

```python
from peft import LoraConfig, get_peft_model, set_peft_model_state_dict
from safetensors.torch import load_file

state = load_file("qwen3-vl-8b/query_lora.safetensors")
targets = sorted({k.split("base_model.model.")[1].rsplit(".lora_", 1)[0] for k in state})
model = get_peft_model(base_model, LoraConfig(r=32, lora_alpha=32, target_modules=targets))
set_peft_model_state_dict(model, state)
```

The files are PEFT LoRA state dicts (keys `base_model.model.<module>.lora_A.weight` and `lora_B.weight`) in float32.