In this video, I’ll be telling you about OpenAI’s newly launched open-weights models – the GPT-OSS 120B and 20B variants. I’ll cover their specifications, capabilities, and how they perform in my updated benchmarks compared to other open models like GLM 4.5 Air.
—
Key Takeaways:
🚀 OpenAI finally released their first major open-weights models – GPT-OSS 117b and 21b with mixture-of-experts architecture.
💻 The 120B model fits on a single 80GB GPU while the 20B runs on just 16GB, making it accessible for consumer hardware.
🧠 These are reasoning-focused models with adjustable effort levels (low, medium, high) and full reasoning process transparency.
📊 In benchmarks, GPT-OSS performs near or below O4-mini levels, which is expected for open-weights models of this size.
⚡ Multiple providers like Groq ($0.15-$0.75) and Cerebras ($0.25-$0.69) offer incredibly cheap access to these models.
🔧 Available through Ollama, LM Studio, VLLM, and a free web interface via GPT OSS site with HuggingFace login.
💔 Disappointing performance on coding tasks, especially ThreeJS, rendering, and complex programming challenges.
🆚 GLM 4.5 Air significantly outperforms GPT-OSS on practical coding tasks while running on similar hardware requirements.
📝 Models come with Apache license for maximum permissiveness and include tool-calling capabilities.
—
Timestamps:
00:00 – Introduction to GPT-OSS 120B & 20B
04:49 – KingBench Results
07:11 – GLM 4.5 Air & GPT-OSS 120B Comparison
10:17 – Ending
source
