---
title: Qwythos-27B-v1-OptiQ-4bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/Qwythos-27B-v1-OptiQ-4bit"
md_url: "https://www.modelscope.cn/models/mlx-community/Qwythos-27B-v1-OptiQ-4bit.md"
repository: mlx-community/Qwythos-27B-v1-OptiQ-4bit
last_updated: 2026-10-30
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - qwen3_5
architectures:
  - Qwen3_5ForConditionalGeneration
base_model:
  - empero-ai/Qwythos-27B-v1
base_model_relation: quantized
parameters: 5.5B
tensor_type:
  - BF16
  - U32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 90
stars: 4
tags:
  - mlx
  - optiq
  - quantized
  - 4bit
  - mixed-precision
  - qwen3.5
  - reasoning
  - multimodal
  - vision
  - image-text-to-text
  - long-context
  - 1M-context
  - mtp
  - multi-token-prediction
  - function-calling
  - tool-use
  - agentic
  - apple-silicon
---

# Qwythos-27B-v1-OptiQ-4bit

> Qwythos-27B-v1-OptiQ-4bit - mlx-community 在 ModelScope 开源的模型。mlx-community/Qwythos-27B-v1-OptiQ-4bit

mlx-community/Qwythos-27B-v1-OptiQ-4bit 是 ModelScope 魔搭社区上的 5.5B 参数image-text-to-text模型，采用 apache-2.0 许可，基于 empero-ai/Qwythos-27B-v1 构建。

- **Repository**: mlx-community/Qwythos-27B-v1-OptiQ-4bit
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 5.5B
- **Base model**: empero-ai/Qwythos-27B-v1
- **Tags**: mlx, optiq, quantized, 4bit, mixed-precision, qwen3.5, reasoning, multimodal, vision, image-text-to-text, long-context, 1M-context, mtp, multi-token-prediction, function-calling, tool-use, agentic, apple-silicon
- **Downloads**: 90
- **Stars**: 4
- **Last updated**: 2026-10-30

Source: https://www.modelscope.cn/models/mlx-community/Qwythos-27B-v1-OptiQ-4bit

---

# mlx-community/Qwythos-27B-v1-OptiQ-4bit

> **Built with [mlx-optiq](https://mlx-optiq.com)**, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. [All OptiQ quants](https://mlx-optiq.com/models) · [Docs](https://mlx-optiq.com/docs/)

OptiQ mixed-precision quant of [empero-ai/Qwythos-27B-v1](https://huggingface.co/empero-ai/Qwythos-27B-v1), a Qwen3.5-family vision-language model (image + text, long context, with a bundled MTP speculation head). 19 GB on disk.

## What it is

| Property | Value |
|---|---|
| Base | empero-ai/Qwythos-27B-v1 (Qwen3.5-VL, 64-layer hybrid linear + full attention) |
| Method | OptiQ mixed-precision, per-layer 4/8-bit |
| Bit allocation | Reused from the base [Qwen3.5-27B OptiQ recipe](https://huggingface.co/mlx-community/Qwen3.5-27B-OptiQ-4bit): the architecture is identical, so the per-layer sensitivity ranking transfers directly and no per-model sensitivity sweep is needed |
| Layer split | 279 layers at 4-bit, 219 at 8-bit |
| Achieved bits-per-weight | 5.55 |
| On disk | 19 GB |
| MTP | Speculation head preserved in `optiq/mtp.safetensors` for faster decode via `optiq serve --mtp` |
| Vision | bf16 vision tower kept in `optiq/optiq_vision.safetensors` for image and video-frame input |

Following the naming `llama.cpp` uses for its mixed quants, the "4bit" label denotes the family, not the weighted average.

## Run it

Qwen3.5 plus the MTP and vision sidecars register through OptiQ, so `import optiq` once before loading:

```bash
pip install "mlx-optiq>=0.4.7"
```

```python
import optiq  # registers the arch + MTP/vision sidecars
from mlx_lm import load, generate

model, tok = load("mlx-community/Qwythos-27B-v1-OptiQ-4bit")
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Explain mixed-precision quantization in two sentences."}],
    tokenize=False, add_generation_prompt=True,
)
print(generate(model, tok, prompt=prompt, max_tokens=400))
```

For image and video input, plus an OpenAI + Anthropic-compatible endpoint with the MTP speculation head and mixed-precision KV cache:

```bash
optiq serve --model mlx-community/Qwythos-27B-v1-OptiQ-4bit
```

Qwythos is a reasoning model, so give it a generous token budget.

Using the head outside `optiq serve`: `optiq/mtp.safetensors` holds Qwen's original MTP tensors, so its seven RMSNorm weights (`mtp.*norm*`) are zero-centred, applied as 1 + w. The main weights have the +1 folded in, as MLX expects. Add 1.0 to all seven norms when loading the head into another runtime.

## Links

- **Project website:** [mlx-optiq.com](https://mlx-optiq.com/)
- **All OptiQ quants:** [mlx-optiq.com/models](https://mlx-optiq.com/models)
- **Base model:** [empero-ai/Qwythos-27B-v1](https://huggingface.co/empero-ai/Qwythos-27B-v1)