Remove throughput and efficiency tables

README.md

@@ -111,27 +111,6 @@ TEG's image-text retrieval trades accuracy for edge deployability — MobileNetV
 | Nemotron-3B | 4.7B | 0.727 |
 | ImageBind | 1.2B | 0.664 |
 
-### Throughput — items/s on NVIDIA L4
-
-| Model | Params | Text | Audio | Image |
-|---|---|---|---|---|
-| EBind | ~200M | **771** | 51.0 | 46.7 |
-| **TEG-421M** | **421M** | 470 | **180** | **158** |
-| ImageBind | 1.2B | 122 | 49.5 | 20.6 |
-| LCO-3B | 4.7B | 90 | 5.2 | 15.4 |
-| Nemotron-3B | 4.7B | 90 | 42.8 | 15.4 |
-
-TEG is **35x faster than LCO on audio** and **10x faster on image**. It leads all models on audio and image throughput while being the only one with both edge-deployable encoders and competitive retrieval quality.
-
-### Efficiency comparison
-
-| Model | Params | SALT T→A R@1 | Audio items/s | R@1 per billion params |
-|---|---|---|---|---|
-| **TEG-421M** | 0.42B | **0.117** | **180** | **0.278** |
-| EBind | ~0.2B | 0.047 | 51.0 | 0.235 |
-| ImageBind | 1.2B | 0.038 | 49.5 | 0.032 |
-| LCO-3B | 4.7B | 0.042 | 5.2 | 0.009 |
-| Nemotron-3B | 4.7B | 0.018 | 42.8 | 0.004 |
 
 ## Usage