SINGHA Tiny VLM v2
A 561,370-parameter, 37.9M-MAC vision model that outputs 194 slot logits (objects, main subject, count, scene, action, confidence gates), turned into a short caption by a tiny rule-based captioner. Built for the ESP32-S3 (INT8, ESP-DL).
- Input: RGB 128x128, x = pixel/127.5 - 1, NCHW
- Files: singha_tiny_vlm_128.onnx, singha_tiny_vlm_esp32s3_128_w8a8.espdl, singha_caption_spec.json, singha_caption_spec.h, singha_final.pt, random_test_grid.png
- Training: supervised distillation from YOLOv8m + CLIP, then judge-driven RL (Qwen2-VL-2B)
- Shipped model: rl
- Accuracy is agreement with teacher labels, not human ground truth. On-board latency is not measured.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support