Henry Ndubuaku commited on
Drop the Performance row
Browse files
README.md
CHANGED
|
@@ -35,7 +35,6 @@ Needle does three jobs, all of them on the device:
|
|
| 35 |
| Outputs | Tool calls or structured extraction |
|
| 36 |
| Model | 29-121M Laddered Simple Attention Networks, CQ2 quantisation |
|
| 37 |
| Training | 360B tokens of proprietary structured dataset |
|
| 38 |
-
| Performance | Competes with LFM2.5-1.2B, Qwen3.5-0.8B and AFM3 |
|
| 39 |
| Speed | 400-4k tokens/s decode and 1-10k tokens/s prefill on a Raspberry Pi 5 |
|
| 40 |
| Customisation | Fine-tuning lifts every subnetwork 18 to 36 points on DroidCall; from 4 layers (29M) up the tuned subnetwork passes DeepSeek V4 Flash |
|
| 41 |
|
|
|
|
| 35 |
| Outputs | Tool calls or structured extraction |
|
| 36 |
| Model | 29-121M Laddered Simple Attention Networks, CQ2 quantisation |
|
| 37 |
| Training | 360B tokens of proprietary structured dataset |
|
|
|
|
| 38 |
| Speed | 400-4k tokens/s decode and 1-10k tokens/s prefill on a Raspberry Pi 5 |
|
| 39 |
| Customisation | Fine-tuning lifts every subnetwork 18 to 36 points on DroidCall; from 4 layers (29M) up the tuned subnetwork passes DeepSeek V4 Flash |
|
| 40 |
|