Alibaba Qwen · First MLX build 28 Apr 2025 · In Core AI since 8 Jun 2026
Qwen3-4B
MLXCore AI
Text apache-2.0
A 4B parameter text generation model from Alibaba Qwen, with 8 MLX builds, from 3-bit to bf16 and Apple Core AI export recipes for macOS and iOS.
- Parameters
- 4B From Hugging Face
- Quantisation
- 4-bit Qwen3-4B-4bit
- Memory needed
- 5 GB Estimate, unified memory
- Download
- 2.3 GB Weights, estimated
- Downloads
- 30K This build, last 30 days
- Task
- Text generation Text
- MLX builds
- 8 On mlx-community
- Devices
- Mac Apple silicon
- Parameters
- 4B From Hugging Face
- Compression
- 4-bit, 4-bit weights, 8-bit KV cache, None Apple’s preset
- Memory needed
- 5 GB Estimate from parameters
- Context
- 40K tokens Fixed at export
- Precision
- FP16 Compute precision
- Platforms
- macOS, iOS Mac and iPhone
- Export presets
- 3 In Apple’s registry
- Listed by Apple
- 8 Jun 2026 First registry commit
Memory
Will it fit?Needs about 5 GB at 4-bit
8 GB
16 GB
24 GB
32 GB
64 GB
128 GB
Fits with headroomTight, close other appsWon’t fit
Estimated from the size of the weights, plus a fifth for context and working memory and 2 GB for macOS and your apps. Long contexts need more.
Run it
Terminal · mlx-lm$ pip install -U mlx-lm
$ mlx_lm.generate --model mlx-community/Qwen3-4B-4bit --prompt "Hello"Memory
Will it fit?Needs about 5 GB at 4-bit, 4-bit weights, 8-bit KV cache, None
8 GB
16 GB
24 GB
32 GB
64 GB
128 GB
Fits with headroomTight, close other appsWon’t fit
Estimated from the parameter count at Apple’s compression setting. Apple does not publish package sizes.
Export it
Terminal · uv$ git clone https://github.com/apple/coreai-models
$ cd coreai-models
$ uv run coreai.llm.export qwen3-4b
$ uv run coreai.llm.export qwen3-4b-8bit-kv
$ uv run coreai.llm.export qwen3-4b --platform iOS
Apple publishes the recipe, not a ready-made package: these commands convert the original weights to a Core AI .aimodel on your Mac. See the repository for requirements.
MLX builds
8 on mlx-community| Qwen3-4B-3bit | 3-bit | 1.8 GB | 4 GB | 170 |
|---|---|---|---|---|
| Qwen3-4B-4bit | 4-bit | 2.3 GB | 5 GB | 30K |
| Qwen3-4B-4bit-DWQ | 4-bit DWQ | 2.3 GB | 5 GB | 173 |
| Qwen3-4B-4bit-DWQ-053125 | 4-bit DWQ | 2.3 GB | 5 GB | 145 |
| Qwen3-4B-4bit-DWQ-calibration-1332 | 4-bit DWQ | 2.3 GB | 5 GB | 33 |
| Qwen3-4B-6bit | 6-bit | 3.3 GB | 6 GB | 83 |
| Qwen3-4B-8bit | 8-bit | 4.3 GB | 7 GB | 1.3K |
| Qwen3-4B-bf16 | bf16 | 8.0 GB | 11 GB | 241 |
Sizes are estimated from Hugging Face’s parameter counts and the quantisation. Downloads are for the last 30 days.
Core AI export presets
From Apple’s registry| Preset | Platform | Compression | Precision | Context |
|---|---|---|---|---|
| qwen3-4b | macOS | 4-bit | FP16 | 40K |
| qwen3-4b-8bit-kv INT4 per-block weights and INT8 per-tensor KV cache. |
macOS | 4-bit weights, 8-bit KV cache | FP16 | 40K |
| qwen3-4b | iOS | None | FP16 | 4K |