06/08/2025
🚨 BREAKING: OpenAI just went OPEN
For the first time since GPT‑2, OpenAI released open-weight models you can actually download, run, and fine-tune.
They’re called gpt-oss-20b and gpt-oss-120b, and here’s why it matters:
🧠 What are they?
OpenAI dropped two text-only models built for reasoning and tool use:
• 🧱 gpt-oss-20b
• 🧠 gpt-oss-120b
"Open-weight" means the trained model weights are public and free to use under the Apache 2.0 license.
You can run them, fine-tune them, and use them commercially.
But the training data and code remain private.
🚀 Performance
• gpt-oss-120b matches or beats o4-mini on reasoning tasks like coding, competitive math, and medical queries
• gpt-oss-20b outperforms o3-mini on math and health benchmarks
Both support tool use like web search, code ex*****on, and agents, but they’re not multimodal (no image generation).
📦 Architecture
Both models use a Mixture-of-Experts (MoE) design:
• 🧠 gpt-oss-120b: 117B total parameters, ~5.1B active per token
• 🧱 gpt-oss-20b: 21B total, ~3.6B active
They support 128,000-token context windows, long enough to fit an entire book.
🖥️ Hardware Requirements
• 🧱 gpt-oss-20b runs on laptops or consumer GPUs with 16GB VRAM or Apple Silicon
• 🧠 gpt-oss-120b needs 60–80GB VRAM or large-memory cloud GPUs
🧩 Experts per Layer
Leaked details confirm:
• gpt-oss-120b uses 128 experts per layer, with 4 experts activated per token.
This helps it stay efficient while delivering top-tier reasoning.
🔐 Licensing & Safety
• Released under Apache 2.0, commercial use allowed
• Tested for adversarial risks, results within safe limits
• Sensitive data filtered out (e.g., biological hazards)
☁️ Where to Run It
Models are available on:
• Hugging Face
• GitHub
• AWS (via Bedrock & SageMaker)
• Integrated with Ollama, vLLM, Microsoft AI Toolkit
You can run it locally or scale it in the cloud.