---
base_model_relation: quantized
language: en
library_name: mlx
pipeline_tag: image-text-to-text
tags:
- mlx
- text-generation-inference
- mlx-vlm
- mimo_v2
- agentic
- distillation
- supervised-fine-tuning
- code
- tool-use
- apple-silicon
- 4-bit
- 8-bit
license: other
base_model:
- XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
---

# **MiMo-V2.6-Distill-Qwen-9B-MLX**

> **MiMo-V2.6-Distill-Qwen-9B** is a 9-billion-parameter agentic model from Xiaomi MiMo, built via supervised fine-tuning of Qwen3.5-9B on MiMo-generated data spanning coding, general-purpose agent tasks, visual coding, and cybersecurity, released as an SFT checkpoint intended as a starting point for open research in agentic reinforcement learning. It was trained on a weighted mixture of 77.4B total tokens (27.2B loss-bearing) split roughly evenly across code (29.9%), general (28.5%), visual (27.4%), and cyber (14.2%) domains, and shows dramatic gains over its Qwen3.5-9B base across nearly every reported benchmark — jumping from 5.0 to 30.3 on AutomationBench v1.0.6, 5.7 to 31.3 on internal cyber evaluation, 27.0 to 37.1 on Terminal-Bench 2.1, and 19.5 to 51.6 on internal coding evaluation, alongside solid improvements on SWE-bench Verified (61.1) and SWE-bench Pro (44.6). It ships with its tokenizer and MiMo v2.6 chat template, is served via SGLang with a dedicated `mimo` reasoning parser and explicit thinking-mode control through an OpenAI-compatible API, and is fine-tuned from the Apache-2.0-licensed Qwen3.5-9B base.

```text
prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX (main)  [~32.83 GB Total]
├── 4bit/                                          [~5.59 GB]
│   ├── model shards: 2 safetensors (~5.56 GB)
│   └── metadata & configs (~26.7 MB)
│
├── 8bit/                                          [~9.73 GB]
│   ├── model shards: 2 safetensors (~9.70 GB)
│   └── metadata & configs (~26.7 MB)
│
└── [BF16 Base Files]                              [~17.53 GB]
    ├── model shards: 4 safetensors (~17.50 GB)
    └── metadata & configs (~26.7 MB)
```

## Use with mlx

Install the latest version of `mlx-vlm`:

```bash
pip install -U mlx-vlm
```

> **Model Note:** `MiMo-V2.6-Distill-Qwen-9B` is an agentic vision-language model trained for coding, cybersecurity, general multi-modal reasoning, and visual problem-solving. It accepts both text and image inputs.

### BF16 Variant (Base Model)

The unquantized BF16 weights are located directly in the root of the repository:

#### CLI (Terminal)

```bash
python -m mlx_vlm generate \
  --model prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX \
  --max-tokens 512 \
  --temperature 0.2 \
  --prompt "Analyze this image and explain the core workflow step by step." \
  --image <path_to_image>
```

#### Python API

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX"
model, processor = load(model_path)
config = load_config(model_path)

image = ["<path_to_image>"]
prompt = "Analyze this image and explain the core workflow step by step."
formatted_prompt = apply_chat_template(processor, config, prompt, num_images=len(image))

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    image=image, 
    max_tokens=512, 
    temperature=0.2
)
print(output.text)
```

### 8-bit Variant

Target the `8bit` subfolder:

#### CLI (Terminal)

```bash
python -m mlx_vlm generate \
  --model prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX/8bit \
  --max-tokens 512 \
  --temperature 0.2 \
  --prompt "Analyze this image and explain the core workflow step by step." \
  --image <path_to_image>
```

#### Python API

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX"
model, processor = load(model_path, subfolder="8bit")
config = load_config(model_path, subfolder="8bit")

image = ["<path_to_image>"]
prompt = "Analyze this image and explain the core workflow step by step."
formatted_prompt = apply_chat_template(processor, config, prompt, num_images=len(image))

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    image=image, 
    max_tokens=512, 
    temperature=0.2
)
print(output.text)
```

### 4-bit Variant

Target the `4bit` subfolder:

#### CLI (Terminal)

```bash
python -m mlx_vlm generate \
  --model prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX/4bit \
  --max-tokens 512 \
  --temperature 0.2 \
  --prompt "Analyze this image and explain the core workflow step by step." \
  --image <path_to_image>
```

#### Python API

```python
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/MiMo-V2.6-Distill-Qwen-9B-MLX"
model, processor = load(model_path, subfolder="4bit")
config = load_config(model_path, subfolder="4bit")

image = ["<path_to_image>"]
prompt = "Analyze this image and explain the core workflow step by step."
formatted_prompt = apply_chat_template(processor, config, prompt, num_images=len(image))

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    image=image, 
    max_tokens=512, 
    temperature=0.2
)
print(output.text)
```

## License and Attribution

* **Model:** [XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B)
* **License:** This MLX compression follows the same license as the original [MiMo-V2.6-Distill-Qwen-9B](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B) model.
* **MLX-VLM:** [Blaizzy/mlx-vlm](https://github.com/Blaizzy/mlx-vlm)
* **MLX:** [ml-explore/mlx](https://github.com/ml-explore/mlx)