SuperOptiX now supports OpenAI's latest open-source language models: GPT-OSS-20B and GPT-OSS-120B!
What are GPT-OSS Models?
OpenAI has released two powerful open-weight language models designed for advanced reasoning and complex tasks. The GPT-OSS-120B model contains 117B parameters with 5.1B active parameters, making it suitable for production environments and high-reasoning use cases. The GPT-OSS-20B model contains 21B parameters with 3.6B active parameters, optimized for lower latency and specialized use cases.
Key Features
These models come with several advanced capabilities:
Permissive Apache 2.0 License: Build freely without copyleft restrictions or patent risk. This makes these models ideal for experimentation, customization, and commercial deployment.
Configurable Reasoning Effort: Adjust reasoning effort (low, medium, high) based on your specific use case and latency needs. This flexibility allows you to optimize performance based on your application requirements.
Full Chain-of-Thought: Gain complete access to the model's reasoning process, facilitating easier debugging and increased trust in outputs. This feature is particularly valuable for developers who need to understand how the model arrives at its conclusions.
Fine-tunable: Fully customize models to your specific use case through parameter fine-tuning. The larger GPT-OSS-120B model can be fine-tuned on a single H100 node, while the smaller GPT-OSS-20B model can even be fine-tuned on consumer hardware.
Agentic Capabilities: Use the models' native capabilities for function calling, web browsing, Python code execution, and Structured Outputs. These capabilities enable the models to interact with external systems and perform complex multi-step tasks.
Native MXFP4 Quantization: The models are trained with native MXFP4 precision for the MoE layer, making GPT-OSS-120B run on a single H100 GPU and GPT-OSS-20B fit within 16GB of memory.
SuperOptiX Integration
SuperOptiX, the comprehensive AI agent development framework built on DSPy, now provides seamless integration with GPT-OSS models. While SuperOptiX currently supports basic model management and inference, we're actively working on implementing the advanced features mentioned above. These capabilities will be available in future releases.
How to Use GPT-OSS with SuperOptiX
You can install GPT-OSS models via Ollama, MLX, or HuggingFace.
π¦ Recommended: Use Ollama for GPT-OSS Models
For the best performance and reliability with GPT-OSS models, we recommend using Ollama:
- β Best Performance: 19.7 t/s vs 5.2 t/s (MLX) vs N/A (HuggingFace)
- β Cross-Platform: Works on all platforms (Windows, macOS, Linux)
- β Easy Setup: Simple installation and model management
- β Optimized Format: GGUF format optimized for local inference
- β No Server Required: Direct model execution
β οΈ Warning for Apple Silicon Users
GPT-OSS models (openai/gpt-oss-20b, openai/gpt-oss-120b) cannot run on Apple Silicon via HuggingFace backend due to mixed precision conflicts between F16 and BF16 data types that Apple's Metal Performance Shaders doesn't support.
Use these alternatives for full Apple Silicon compatibility:
# Best option: Ollama (cross-platform, best performance)
super model run gpt-oss:20b "prompt" --backend ollama
# Apple Silicon native: MLX (requires server)
super model run lmstudio-community/gpt-oss-20b-MLX-8bit "prompt" --backend mlx
Note: Downloading these models may take a significant amount of time. The GPT-OSS-120B model is approximately 120GB, while the GPT-OSS-20B model is around 40GB. Make sure you have sufficient storage space and a stable internet connection.
Quick Start with Ollama (RECOMMENDED)
You need to update Ollama to get the latest GPT-OSS models.
# Install GPT-OSS models
super model install gpt-oss:20b
super model install gpt-oss:120b
# Test the models
super model run gpt-oss:20b "Explain quantum computing with detailed reasoning"
π Using MLX Backend (Apple Silicon)
For Apple Silicon users who prefer native MLX performance:
# Install GPT-OSS models via MLX
super model install lmstudio-community/gpt-oss-20b-MLX-8bit --backend mlx
super model install lmstudio-community/gpt-oss-120b-MLX-8bit --backend mlx
# Start MLX server
super model server mlx lmstudio-community/gpt-oss-20b-MLX-8bit --port 8000
# Test the models
super model run lmstudio-community/gpt-oss-20b-MLX-8bit "Explain quantum computing" --backend mlx
π€ Using HuggingFace Backend
You need to install the HuggingFace dependencies: pip install superoptix[huggingface]. Then you should be able to download the models directly from HuggingFace.
# Install via HuggingFace
super model install openai/gpt-oss-20b --backend huggingface
super model install openai/gpt-oss-120b --backend huggingface
# Start server
super model server huggingface openai/gpt-oss-20b --port 8001
Configuring Playbooks with GPT-OSS Models
Once you have installed the GPT-OSS models, you can configure your SuperOptiX playbooks to use them. Here are examples of how to set up different types of agents:
Basic Playbook Configuration
To use GPT-OSS models in your SuperOptiX agents, you need to configure the language_model section in your playbook YAML file:
π¦ Ollama Backend (Cross-platform - RECOMMENDED):
language_model:
provider: ollama
model: gpt-oss:20b # or gpt-oss:120b
api_base: http://localhost:11434
temperature: 0.7
max_tokens: 2048
οΏ½οΏ½ MLX Backend (Apple Silicon - Native Support):
language_model:
provider: mlx
model: lmstudio-community/gpt-oss-20b-MLX-8bit
api_base: http://localhost:8000
temperature: 0.7
max_tokens: 2048
π€ HuggingFace Backend (Limited on Apple Silicon):
language_model:
provider: huggingface
model: openai/gpt-oss-20b
api_base: http://localhost:8001
temperature: 0.7
max_tokens: 2048
This configuration tells SuperOptiX to use the specified GPT-OSS model for your agent's language processing tasks. The temperature parameter controls the creativity of the responses, while max_tokens limits the length of generated text.
Performance Comparison
| Backend | Model | Status | Performance | Apple Silicon | Recommendation |
|---|---|---|---|---|---|
| οΏ½οΏ½ Ollama | gpt-oss:20b | β Works | depends | β Optimized format | β RECOMMENDED |
| π MLX-LM | lmstudio-community/gpt-oss-20b-MLX-8bit | β Works | depends | β Native support | Apple Silicon only |
| π€ HuggingFace | openai/gpt-oss-20b | β Broken | N/A | β Mixed precision errors | β Avoid on Apple Silicon |
Check model status
super model list | grep gpt-oss
This process verifies that your GPT-OSS model is properly configured and accessible to your SuperOptiX agent.
Hardware Requirements
GPT-OSS-20B requires approximately 16GB of RAM and is well-suited for local development and testing. GPT-OSS-120B requires an H100 GPU but I managed to fit this in M4 Max 128GB, is designed for production environments that demand the highest level of reasoning capabilities.
Getting Started
- Install SuperOptiX:
pip install superoptix(You can add MLX and Hugging Face as needed) - Install GPT-OSS models:
super model install gpt-oss:20b - Test the model:
super model run gpt-oss:20b "Hello, how are you?" - Create your first agent:
super init my-gpt-oss-project
Future Enhancements
SuperOptiX is actively working on implementing advanced GPT-OSS features including configurable reasoning levels, chain-of-thought integration, agentic capabilities, and fine-tuning support. These capabilities will be available in future releases.
- MLX Support: Work in Progress to support new "gpt-oss" architecture. Coming soon for Apple Silicon optimization
- Advanced Tooling: Enhanced function calling and web browsing
- Performance Monitoring: Detailed analytics and optimization
- Community Models: Support for fine-tuned GPT-OSS variants
Resources
SuperOptiX - Empowering developers to build intelligent AI agents with the latest open-source models.
Top comments (0)