---
language:
- en
license: apache-2.0
base_model:
- kai-os/Carnice-9b
tags:
- qwen3.5
- deltanet
- hipfire
- rdna
- amd
- hermes
- tool-use
pipeline_tag: text-generation
---

# Carnice-9B — hipfire quantized

[kai-os/Carnice-9b](https://huggingface.co/kai-os/Carnice-9b) quantized for [hipfire](https://github.com/Kaden-Schutt/hipfire), a Rust-native inference engine for AMD RDNA GPUs.

Carnice is a Hermes tool-use finetune of Qwen3.5-9B with browser, terminal, and reasoning capabilities.

## Files

| File | Quant | Size | Speed (5700 XT) | Notes |
|------|-------|------|-----------------|-------|
| `carnice-9b.hf4` | HF4 (4-bit) | 4.5 GB | ~41 tok/s | Faster, fits 8GB VRAM |
| `carnice-9b.hf6` | HF6 (6-bit) | 6.9 GB | ~32 tok/s | Better quality, needs 8GB |

## Usage

```bash
# Install hipfire
curl -L https://raw.githubusercontent.com/Kaden-Schutt/hipfire/master/scripts/install.sh | bash

# Pull and run
hipfire pull carnice:9b        # HF4 (default)
hipfire pull carnice:9b-hf6    # HF6 (higher quality)
hipfire run carnice:9b
```

## About

Carnice uses the Hermes chat template with `<tool_call>` XML tags for function calling. hipfire passes these through as regular tokens — the model generates tool calls naturally within the ChatML format.

Same DeltaNet hybrid architecture as base Qwen3.5. JIT-compiles for any AMD GPU (RDNA 1-4, APUs).

## License

Apache 2.0