---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- vLLM
- AWQ
- autoawq
- INT4
- W4A16
- qwen3.5
- reasoning
- long-context
- 1M-context
- function-calling
- tool-use
- cybersecurity
- biomedical
- agentic
base_model:
- empero-ai/Qwythos-9B-Claude-Mythos-5-1M
base_model_relation: quantized
---

# Qwythos-9B-Claude-Mythos-5-1M-AWQ

Base model: [empero-ai/Qwythos-9B-Claude-Mythos-5-1M](https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M)

This repository contains an **AWQ INT4 / W4A16** quantized export of Qwythos-9B-Claude-Mythos-5-1M.

The model was quantized on an **NVIDIA GeForce RTX 4090** with a self-implemented AWQ exporter that writes a vLLM / AutoAWQ-style weight layout:

- `.qweight`
- `.qzeros`
- `.scales`
- sharded `model-xxxxx-of-xxxxx.safetensors`
- `model.safetensors.index.json`

The quantization path is **data-free**: no external calibration dataset was used. The exporter estimates per-layer input importance from the original weights, then searches weighted group clipping / minmax scales to reduce reconstruction error.

This release is primarily prepared for **vLLM** with AWQ GEMM weight-only inference. It does not claim compatibility with every runtime.

## Quantization Policy

Qwythos is based on Qwen3.5, whose hybrid attention stack contains linear-attention and attention modules that should remain in native precision for vLLM loading. For this reason, this export intentionally quantizes only the text MLP projection weights.

| Scope | Format |
|-------|--------|
| Text MLP `gate_proj` | W4A16 AWQ, group size 128, zero point |
| Text MLP `up_proj` | W4A16 AWQ, group size 128, zero point |
| Text MLP `down_proj` | W4A16 AWQ, group size 128, zero point |
| First text layer | Native BF16/FP16 |
| `linear_attn.*` | Native BF16/FP16 |
| `self_attn.*` | Native BF16/FP16 |
| Embeddings, norms, rotary embeddings | Native BF16/FP16 |
| `lm_head` | Native BF16/FP16 |
| Vision / visual modules | Native BF16/FP16 |
| MTP / special heads, if present | Native BF16/FP16 |

Quantization config used by the exporter:

```yaml
w_bit: 4
q_group_size: 128
zero_point: true
awq_grid_size: 8
awq_samples: 8
calibration_mode: data_free
max_shard_size: 3GB
```

The exporter keeps the Qwen3.5 linear-attention tensors in their official split layout rather than fusing them, which is important for vLLM compatibility.

## Dependencies

Recommended vLLM serving stack:

```bash
pip install --upgrade vllm transformers accelerate safetensors
```

Qwen3.5 support in serving frameworks is still relatively new. If your local vLLM build cannot load `qwen3_5`, upgrade vLLM and Transformers first.

## vLLM Startup Command

Start with a conservative context length on a single RTX 4090. The source model advertises a 1M context window, but KV cache memory, not AWQ weight size, becomes the limiting factor at long context.

```bash
vllm serve JunHowie/Qwythos-9B-Claude-Mythos-5-1M-AWQ \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name Qwythos-9B-AWQ \
  --trust-remote-code \
  --quantization awq \
  --dtype float16 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.85
```

If memory headroom is sufficient, increase `--max-model-len` gradually.

## Example Request

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwythos-9B-AWQ",
    "messages": [
      {
        "role": "user",
        "content": "Explain how AWQ quantization protects salient channels, and why Qwen3.5 linear-attention tensors are kept native."
      }
    ],
    "temperature": 0.6,
    "top_p": 0.95,
    "max_tokens": 4096
  }'
```

## Model Download

```python
from huggingface_hub import snapshot_download

snapshot_download(
    "JunHowie/Qwythos-9B-Claude-Mythos-5-1M-AWQ",
    cache_dir="your_local_path",
)
```

## Notes

- This is a quantized derivative of Qwythos-9B-Claude-Mythos-5-1M, not a new fine-tune.
- The original model is a reasoning model and may emit `<think>...</think>` blocks before the final answer.
- Recommended sampling follows the base model card: `temperature=0.6`, `top_p=0.95`, `top_k=20`, `repetition_penalty=1.05`.
- The source model advertises a 1,048,576-token context window with YaRN rope scaling. For this AWQ release, start smaller on consumer GPUs and raise context length after memory testing.
- No benchmark numbers are changed or claimed by this AWQ export. Use the original model card for upstream capability references.

## Quantization Log

```text
2026-06-30
1. Quantized empero-ai/Qwythos-9B-Claude-Mythos-5-1M on RTX 4090.
2. Exported AWQ GEMM W4A16 weights with group size 128 and zero point.
3. Used data-free AWQ search; no calibration dataset was involved.
4. Kept Qwen3.5 attention, linear-attention, embeddings, norms, lm_head, visual modules, and first text layer in native precision for vLLM compatibility.
```

## Original Model Overview

Qwythos-9B is developed by [Empero](https://empero.org). It is a full-parameter reasoning model built on top of a deeply uncensored Qwen3.5-9B base and post-trained on over 500 million tokens of Claude Mythos and Claude Fable traces, with chain-of-thought generated by Empero AI's internal tool `rethink`.

Key upstream features:

- **1,048,576-token context** via YaRN rope scaling enabled by default.
- **Native function calling** following Qwen3.5's tool-use format.
- **Reasoning-oriented behavior** with `<think>` blocks.
- **Strong gains over the base Qwen3.5-9B** in the original evaluation setup, especially on MMLU and GSM8K.
- **Domain emphasis** on cybersecurity, biomedical, quantitative reasoning, coding, and agentic workflows.

Original evaluation headline from the base model card:

| Task | Metric | Base Qwen3.5-9B | Qwythos-9B | Delta |
|---|---|---:|---:|---:|
| gsm8k | exact_match flexible | 0.670 | 0.860 | +0.190 |
| gsm8k | exact_match strict | 0.510 | 0.810 | +0.300 |
| mmlu | acc | 0.232 | 0.575 | +0.343 |
| arc_challenge | acc | 0.470 | 0.490 | +0.020 |
| arc_challenge | acc_norm | 0.400 | 0.410 | +0.010 |
| gpqa_diamond | exact_match flexible | 0.630 | 0.580 | -0.050 |

These numbers belong to the original FP/BF16 model card and are included only as upstream reference.

## Base Model Details

- **Upstream model:** [empero-ai/Qwythos-9B-Claude-Mythos-5-1M](https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M)
- **Underlying base:** [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)
- **Fine-tune type:** full-parameter supervised fine-tune
- **Context length:** 1,048,576 tokens with YaRN rope scaling
- **License:** Apache-2.0

## Limitations

- This AWQ export has not been presented as a full formal benchmark release.
- Very long context still requires substantial KV-cache memory.
- For exact identifiers, safety-critical medical details, live facts, or security-sensitive workflows, pair the model with retrieval/tools and verify outputs.
- Qwythos is intentionally uncensored. Add application-level policy and review layers for public-facing deployments.

## Acknowledgements

- Original model: [Empero Qwythos-9B-Claude-Mythos-5-1M](https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M)
- Underlying base: [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)
- Quantized release: `JunHowie/Qwythos-9B-Claude-Mythos-5-1M-AWQ`