README.md

---
license: apache-2.0
language:
  - multilingual
base_model: Qwen/Qwen3.5-122B-A10B
base_model_relation: quantized
tags:
  - mlx
  - quantization
  - mixed-precision
  - vision
  - multimodal
  - qwen3.5
  - moe
  - 122B
  - apple-silicon
library_name: mlx
pipeline_tag: image-text-to-text
model_type: qwen3_5_moe
model_name: Qwen3.5-122B-A10B-MLX-3.7bit-VL
metadata:
  parameters:
    - 122B
extra_gated_heading: "Qwen3.5-122B-A10B (MLX 3.7-bit VL)"
---

# Qwen3.5-122B-A10B-MLX-3.7bit-VL

Mixed-precision MLX quantization of [Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) — Alibaba's latest MoE model with full vision support preserved in BF16.

- **3.655 BPW** | **52 GB** | Vision preserved (BF16)

## 🚀 Hardware Optimization

This model brings 122B-class multimodal performance to Apple Silicon. By utilizing advanced mixed-precision quantization, we've compressed the model from uniform 4-bit's 65GB down to **52GB** — a **13GB reduction** — while preserving the full vision pipeline at BF16 precision for lossless image understanding.

This optimization unlocks two distinct local inference experiences:

- **64GB Unified Memory (Minimum):** The uniform 4-bit quantization weighs 65GB and simply cannot fit in 64GB at all. This quantization breaks that barrier — fitting a full 122B multimodal model into 64GB for the first time, pushing the hardware boundaries to make local 122B vision+language inference possible on edge devices.
- **96GB+ Unified Memory (Recommended):** Delivers an uncompromised, buttery-smooth multimodal experience. The efficient footprint frees up massive headroom for the KV cache, completely unlocking ultimate long-context capabilities for both text and vision tasks.

## Quantization

4-tier mixed precision by functional sensitivity:

| Bits | Layers | % Params | Description |
|------|--------|----------|-------------|
| BF16 | — | ~2% | Vision tower, norm, router, conv1d — preserving full visual fidelity |
| 6-bit | — | ~8% | Embeddings, v/o_proj, edge layers, full_attention q/k |
| 4-bit | — | ~3% | DeltaNet attention, shared expert |
| 3-bit | — | ~87% | Expert FFN (256 experts, 8 active/token) |

## Benchmark (M2 Max 96GB)

| | This (3.7bit) | Uniform 4bit |
|--|---------------|-------------|
| Model size | **52 GB** | 65 GB |
| Peak memory (ctx=4k) | **55.1 GB** | 67.7 GB |
| Prefill (1k ctx) | 219.0 tok/s | 210.5 tok/s |
| Prefill (4k ctx) | 230.6 tok/s | 227.7 tok/s |
| Generation (1k ctx) | 36.2 tok/s | 38.3 tok/s |
| Generation (4k ctx) | 33.6 tok/s | 35.8 tok/s |

> Speed is virtually identical, but the 13GB saved makes the difference between a usable and unusable long-context experience on 96GB machines.

## Quality (WikiText-2 Perplexity)

Lower is better. Evaluated on 128 sequences × 2048 tokens.

| Metric | Value |
|--------|-------|
| Mean Perplexity | 5.3536 |
| Median Perplexity | 5.3639 |
| Trimmed Mean Perplexity | 5.6631 |

## Usage

```python
from mlx_vlm import load, generate

model, processor = load("MoringLabs/Qwen3.5-122B-A10B-MLX-3.7bit-VL")

# 文本对话
response = generate(model, processor, prompt="Hello!", max_tokens=200)

# 图像理解
response = generate(model, processor, prompt="Describe this image", image="photo.jpg", max_tokens=200)
print(response)
```

## License

Apache 2.0