---
license: apache-2.0
base_model: Qwen/Qwen3.5-2B-Base
library_name: basert
pipeline_tag: text-generation
tags:
  - basert
  - apple-silicon
  - quantized
  - multimodal
---

# Qwen3.5-2B-Base

[BaseRT](https://github.com/basecompute/baseRT) `.base` builds of [`Qwen/Qwen3.5-2B-Base`](https://huggingface.co/Qwen/Qwen3.5-2B-Base) for fast local inference on Apple Silicon (Metal).

This is the **pretrained base model** (no instruction tuning) — use it for completion-style prompting, research, or fine-tuning baselines. For chat, use [`basecompute/Qwen3.5-2B`](https://huggingface.co/basecompute/Qwen3.5-2B). Hybrid attention (Gated DeltaNet + periodic full attention) with a native vision tower; image inputs work via completion prompts.

## Files

| File | Precision | Size |
|------|-----------|------|
| `Qwen3.5-2B-Base-Q4.base` | 4-bit | 1.3 GB |
| `Qwen3.5-2B-Base-Q8.base` | 8-bit | 2.3 GB |

## Usage

```bash
curl -LsSf https://basecompute.co/install.sh | sh
basert pull basecompute/Qwen3.5-2B-Base
basert complete basecompute/Qwen3.5-2B-Base --prompt "The capital of France is"
```

Released under the apache-2.0 license, inherited from the base model.