Sieve-Vision-4B
Model card for Sieve-Vision-4B, a vision decision model for multiple-choice questions about images. It returns probabilities and does not generate text.
Hugging Face: sthanika-ai/Sieve-Vision-4B (Visual Question Answering, about 5B on the listing)
API model name: not available yet. Images are not supported by the API for now.
Sieve-Vision-4B answers multiple-choice questions about one or more images. It returns a probability for each option. It never generates text or captions.
Key facts
| Item | Value |
|---|---|
| Base model | Qwen/Qwen3.5-4B-Base, revision 1001bb4d, 4.66B parameters |
| Trained weights | Full fine-tune of the 32 text decoder layers (3.57B parameters) plus a new 1.57M-parameter decision head. Vision encoder, merger and token embeddings unchanged. |
| Architecture | 24-layer vision encoder, 16 px patches, 2 x 2 merging (one token per 32 x 32 px). Text model: 32 layers, 24 Gated DeltaNet and 8 full-attention. |
| Question type | choice only. Any number of options. One or more images. |
| Memory | 9.3 GB for bf16 weights, plus activations |
Limits
- Images: up to 1,638,400 px (1.6 MP) per image at native resolution. Larger images are downscaled with LANCZOS, keeping aspect ratio.
- Training used 1 to 10 images per question.
- The card does not state a maximum input or output token count.
- Trained on English questions only.
Calibration
- One global temperature, T = 1.817, fitted on the dev split.
- Temperature changes probabilities, not the winning option.
Benchmarks (self-reported)
- Public vision Decision Index: 64.95 on the 0.3 public suite. Not yet on the leaderboard.
- Own held-out test (1,142 questions): 89.8% accuracy, 3.0% ECE after temperature scaling.
- Weak areas: expert-level reasoning (R-Bench-M, MMMU-Pro). Hateful Memes skill 22.2.
Latency (self-reported)
778 ms median and 781 ms p80 per request, with 4 images at 1.6 MP each, on one A100 80GB, batch size 1.
Restrictions
- Do not use this model for content moderation. It was not trained for that task and scores 22.2 on Hateful Memes.
- It does not generate captions or free text.
Requirements
torch>=2.8,transformers>=5.16,safetensors,pillow,huggingface_hub.flash-linear-attention(recommended; the authors used 0.5.2).
Licence
Apache-2.0, inherited from the backbone. The card says the LICENSE file is the backbone's. The backbone terms are pending legal review.
Data retention
The model keeps nothing between calls. Through the hosted API, Sieve does not store prompts, state or answers (see data retention). The rules for images will be stated when the API supports them.
Call
Not available through the API yet. When images go live, the request uses the TypeSafe-Danger-Zone-Images: true header and an assets map. The image format and limits will be confirmed when the API supports images.
Next steps
Last updated