Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- mlx
|
| 7 |
+
- apple-silicon
|
| 8 |
+
- multimodal
|
| 9 |
+
- vision-language
|
| 10 |
+
- pixtral
|
| 11 |
+
- llava
|
| 12 |
+
- quantized
|
| 13 |
+
- 3bit
|
| 14 |
+
- 4bit
|
| 15 |
+
- 5bit
|
| 16 |
+
- 6bit
|
| 17 |
+
pipeline_tag: image-text-to-text
|
| 18 |
+
library_name: mlx
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# Apriel-1.5-15B-Thinker — **MLX 3-bit** (Apple Silicon)
|
| 22 |
+
|
| 23 |
+
**Format:** MLX (Mac, Apple Silicon)
|
| 24 |
+
**Quantization:** **3-bit** (balanced footprint ↔ quality)
|
| 25 |
+
**Base:** ServiceNow-AI/Apriel-1.5-15B-Thinker
|
| 26 |
+
**Architecture:** Pixtral-style LLaVA (vision encoder → 2-layer projector → decoder)
|
| 27 |
+
|
| 28 |
+
This repository provides a **3-bit MLX** build of Apriel-1.5-15B-Thinker for **on-device** multimodal inference on Apple Silicon. In side-by-side tests, the **3-bit** variant often:
|
| 29 |
+
- uses **significantly less RAM** than 6-bit,
|
| 30 |
+
- decodes **faster**, and
|
| 31 |
+
- tends to produce **more direct answers** (less “thinking out loud”) at low temperature.
|
| 32 |
+
|
| 33 |
+
If RAM allows, we also suggest trying **4-bit/5-bit/6-bit** variants (guidance below) for tasks that demand more fidelity.
|
| 34 |
+
|
| 35 |
+
> Explore other Apriel MLX variants under the `mlx-community` namespace on the Hub.
|
| 36 |
+
|
| 37 |
+
---
|
| 38 |
+
|
| 39 |
+
## 🔎 Upstream → MLX summary
|
| 40 |
+
|
| 41 |
+
Apriel-1.5-15B-Thinker is a multimodal reasoning VLM built via **depth upscaling**, **two-stage multimodal continual pretraining**, and **SFT with explicit reasoning traces** (math, coding, science, tool-use).
|
| 42 |
+
This MLX release converts the upstream checkpoint with **3-bit** quantization for smaller memory and quick startup on macOS.
|
| 43 |
+
|
| 44 |
+
---
|
| 45 |
+
|
| 46 |
+
## 📦 Contents
|
| 47 |
+
|
| 48 |
+
- `config.json` (MLX config for Pixtral-style VLM)
|
| 49 |
+
- `mlx_model*.safetensors` (3-bit shards)
|
| 50 |
+
- `tokenizer.json`, `tokenizer_config.json`
|
| 51 |
+
- `processor_config.json` / `image_processor.json`
|
| 52 |
+
- `model_index.json` and metadata
|
| 53 |
+
|
| 54 |
+
---
|
| 55 |
+
|
| 56 |
+
## 🚀 Quickstart (CLI)
|
| 57 |
+
|
| 58 |
+
**Single image caption**
|
| 59 |
+
```bash
|
| 60 |
+
python -m mlx_vlm.generate \
|
| 61 |
+
--model <this-repo-id> \
|
| 62 |
+
--image /path/to/image.jpg \
|
| 63 |
+
--prompt "Describe this image in two concise sentences." \
|
| 64 |
+
--max-tokens 128 --temperature 0.0 --device mps --seed 0
|