strands-mlx
Running Strands Agents locally on Apple Silicon - inference, fine-tuning, vision in Python
MLX provider for Strands Agents with LoRA training pipelines.
Requirements: Python ≤3.13, macOS/Linux
# Text models
pip install strands-mlx
# With vision/audio/video
pip install "strands-mlx[vision]"
# Environment setup
uv venv --python 3.13 && source .venv/bin/activate
# Install dependencies
uv pip install strands-agents strands-mlx strands-agents-tools
Quick Start
agent.py
from strands import Agent
from strands_mlx import MLXModel
from strands_tools import calculator
model = MLXModel(model_id="mlx-community/Qwen3-1.7B-4bit")
agent = Agent(model=model, tools=[calculator])
agent("What is 29 * 42?")
# Run with uv
uv run agent.py
Architecture
graph LR
A[Agent Conversations] -->|MLXSessionManager| B[Training Data JSONL]
B -->|dataset_splitter| C[train/valid/test]
C -->|mlx_trainer| D[LoRA Adapter]
D -->|MLXModel| E[Domain Expert Agent]
E -.->|Continuous Learning| A
style A fill:#e1f5ff
style E fill:#d4edda
style D fill:#fff3cd
The complete training cycle: Agents collect their own training data → fine-tune themselves → become domain experts → continue learning.
Train Your Own Model
4 steps: Collect → Split → Train → Use
1. Collect Training Data
from strands import Agent
from strands_tools import calculator
from strands_mlx import MLXSessionManager, dataset_splitter, mlx_trainer
session = MLXSessionManager(session_id="my_training", storage_dir="./dataset")
agent = Agent(model=model, tools=[calculator, dataset_splitter, mlx_trainer], session_manager=session)
# Have conversations - auto-saved to JSONL
agent("Teach me about quantum computing")
agent("Calculate 15 * 7")
# Saved to: ./dataset/my_training.jsonl
2. Split Dataset
agent.tool.dataset_splitter(
input_path="./dataset/my_training.jsonl"
)
# Creates train.jsonl, valid.jsonl, test.jsonl (80/10/10 split)
3. Train with LoRA
agent.tool.mlx_trainer(
action="train",
config={
"model": "mlx-community/Qwen3-1.7B-4bit",
"data": "./dataset/my_training",
"adapter_path": "./adapter",
"iters": 200,
"learning_rate": 1e-5,
"batch_size": 1
}
)
4. Use Trained Model
from strands import Agent
from strands_mlx import MLXModel
trained = MLXModel("mlx-community/Qwen3-1.7B-4bit", adapter_path="./adapter")
agent = Agent(model=trained)
agent("Explain quantum computing") # Uses trained knowledge!
Vision Models
from strands_mlx import MLXVisionModel
model = MLXVisionModel(model_id="mlx-community/Qwen2-VL-2B-Instruct-4bit")
agent = Agent(model=model)
agent("Describe: <image>photo.jpg</image>")
agent("Transcribe: <audio>speech.wav</audio>")
agent("What happens: <video>clip.mp4</video>")
Training Tools
| Tool | Purpose |
|---|---|
mlx_trainer |
Background LoRA training |
dataset_splitter |
Split JSONL → train/valid/test |
validate_training_data |
Check format & token counts |
mlx_invoke |
Runtime model switching |
mlx_vision_invoke |
Vision as a tool |
Advanced Training
YAML config file:
model: mlx-community/Qwen3-1.7B-4bit
data: ./training_data
iters: 1000
learning_rate: 1e-5
lora_parameters:
rank: 8
scale: 16.0
lr_schedule:
name: cosine_decay
warmup: 100
optimizer: adamw
Use config:
agent.tool.mlx_trainer(action="train", config="./lora_config.yaml")
Popular Models
Text:
mlx-community/Qwen3-1.7B-4bit(recommended)mlx-community/Qwen3-4B-4bitmlx-community/Llama-3.2-1B-4bitmlx-community/gemma-2-2b-it-4bit
Vision:
mlx-community/Qwen2-VL-2B-Instruct-4bit(recommended)mlx-community/Qwen2-Audio-7B-Instruct(audio)mlx-community/llava-v1.6-mistral-7b-4bit
Community models at mlx-community
Troubleshooting
Out of memory:
config = {
"grad_checkpoint": True,
"batch_size": 1,
"max_seq_length": 1024
}
Model degraded:
config = {
"iters": 200, # Lower for small datasets
"learning_rate": 1e-5 # Conservative
}
Resources
Citation
@software{strands_mlx2025,
author = {Cagatay Cali},
title = {strands-mlx: MLX Model Provider for Strands Agents},
year = {2025},
url = {https://github.com/cagataycali/strands-mlx}
}
Apache 2 License | Built with MLX, MLX-LM, and Strands Agents
Metadata
Release files for strands-mlx 0.2.6
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| strands_mlx-0.2.6.tar.gz | 42.9 kB | Without provenance | Details for strands_mlx-0.2.6.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| strands_mlx-0.2.6-py3-none-any.whl | Python 3 | none | any | Details for strands_mlx-0.2.6-py3-none-any.whl |
Total release size: 87.5 kB
Release files / strands_mlx-0.2.6.tar.gz
| Download URL | strands_mlx-0.2.6.tar.gz |
|---|---|
| Size | 42.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
4c0b230f26aedf6aef2f49ebfaf7c6cc82b89725bad71452ad81df91e5cfb83a
|
|
BLAKE2b-256 checksum How to use checksums |
b28641584c77def11667cdafd94eed007b2fde57c07cecca23aad4f72d25c7c1
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.13.5
|
Release files / strands_mlx-0.2.6-py3-none-any.whl
| Download URL | strands_mlx-0.2.6-py3-none-any.whl |
|---|---|
| Size | 44.6 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
dc62b12d90e3c8ff2994a91d19d6bface2cd20fb4d23215a8548666faac4c869
|
|
BLAKE2b-256 checksum How to use checksums |
0e1a581860db4547e8db66ca33c4afe379c465bfe0cf4a4969c3a942f4c244ed
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.13.5
|
Security for strands-mlx 0.2.6
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.