---
base_model: RekaAI/reka-flash-3
---

[EXL3](https://github.com/turboderp-org/exllamav3) quantization of [reka-flash-3](https://huggingface.co/RekaAI/reka-flash-3), 4 bits per weight.

### HumanEval (argmax)

| Model                                                                          | Q4   | Q8   | FP16 |
| ------------------------------------------------------------------------------ | ---- | ---- | ---- |
| [reka-flash-3-exl3-3bpw](https://huggingface.co/isogen/reka-flash-3-exl3-3bpw) | 87.8 | 90.2 | 90.9 |
| [reka-flash-3-exl3-4bpw](https://huggingface.co/isogen/reka-flash-3-exl3-4bpw) | 89.0 | 88.4 | 87.2 |