A long-awaited surprise
For the first time in six years, OpenAI is giving people real access to its models.
On August 5, 2025, the company released gpt-oss-120b and gpt-oss-20b.
This is not GPT-5.
But it could be even more important.
Why? Because it’s the first time since GPT-2 that OpenAI has shared models with open weights. That means anyone can download them, run them locally, and even use them for business.
What are these models?
- gpt-oss-120b → A massive reasoning model with 117B parameters. It performs close to OpenAI’s o4-mini.
- gpt-oss-20b → A lighter model with 21B parameters. Runs on consumer GPUs with only 16 GB of memory.
- Both are released under the Apache 2.0 license. Free for personal and commercial use.
Why does this matter?
Until now, OpenAI’s advanced models (GPT-3, GPT-4, GPT-4o) were locked behind paid APIs.
Developers had access to results, not to the models themselves.
Now the rules have changed.
As Wired wrote:
“OpenAI just dropped its first open-weight models in over five years… This represents a shift away from proprietary releases.”
(Wired)
For businesses → more independence.
Researchers → more transparency.
For creators → a personal ChatGPT without subscriptions.
Performance so far
OpenAI says:
- gpt-oss-120b matches o4-mini in reasoning.
- gpt-oss-20b is close to o3-mini, but cheaper to run.
Windows Central tested the smaller model with a UK school exam.
The model showed good reasoning but weak accuracy (9/80).
With more context (32k tokens), it reasoned better but slowed down.
So the takeaway is simple:
These models are powerful, but not perfect.
Safety first
OpenAI also focused on safety:
- Both models were tested with the Preparedness Framework.
- They published a full model card with training details and limitations.
- Risks are lower compared to frontier models like GPT-4o.
This shows OpenAI wants to open AI to the world, but in a responsible way.
Where can you get them?
- Free on Hugging Face.
- Available on AWS Bedrock and SageMaker.
- Integrated in Databricks and Google Cloud.
- Full details in the model card.
That means you can:
- Run a chatbot locally.
- Build apps and automations.
- Integrate AI into business workflows.
As TechRadar explained:
“OpenAI has new smaller open models to take on DeepSeek — and they’ll be available on AWS.”
(TechRadar)
Bigger implications
- Business Insider says this could help the West keep pace with China in the AI race.
(Business Insider) - ITPro notes the low-cost deployment: 20b runs on common hardware, while 120b works on an 80 GB A100 GPU.
(ITPro) - El País points out the 128k context window and efficiency focus.
(El País)
Final thoughts
This is not GPT-5.
But it’s a big step forward.
With gpt-oss, OpenAI is opening the door to developers, researchers, and creators worldwide.
It’s a reminder that AI should not only belong to corporations.
Now, it belongs to anyone ready to build.
Wired — OpenAI just released its first open-weight models since GPT-2
Windows Central — We gave OpenAI’s open-source AI a kid’s test
Business Insider — Why China should take note
TechRadar — OpenAI has new smaller open models to take on DeepSeek
ITPro — OpenAI’s open-weight models explained
Pingback: 5 Hidden Smartphone Features That Make Life Easier