---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - compression
  - quantization
  - gguf
  - bittensor
  - ralph
---

# ralph-qwen3-8b-ternary-r4

Ternary-tier GGUF compression of [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) —
architecture and parameter count unchanged (8,190,735,360 weight-bearing parameters), weights
re-stored at reduced bit-width. Submitted to the ternary bit-tier on
[Bittensor subnet 40](https://taostats.io/subnets/40) (Ralph, model-compression).

## License

Released under **Apache License 2.0** — see `LICENSE`. This is a derivative of Qwen3-8B, itself
Apache-2.0; the original copyright/attribution notices and a description of what changed
(including the third-party quantization tooling used) are preserved in `NOTICE`.

## Model overview

| | |
|---|---|
| Parent | [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (Apache-2.0) |
| Parameters | 8,190,735,360 |
| Architecture | unchanged from parent |
| Format | GGUF |
| Bit tier | ternary |
| File size | 2,430,026,816 bytes (~2.26 GiB) |
| Measured bits/weight | ~2.37 (file size / param count; embedding/output tensors kept at higher precision outside the compressed blocks) |
| Quantization method | llama.cpp, imatrix-calibrated |

No retraining or architectural modification — this is a post-training weight re-storage of the
parent's own weights.