WOW! Yesterday OpenAI released two open-weight models: gpt-oss-120b and gpt-oss-

WOW! Yesterday OpenAI released two open-weight models: gpt-oss-120b and gpt-oss-20b. Roughly 120 and 20 Billion parameters in size and bother featuring reasoning capabilities.

The models utilise quantisation and MoE(mixture of expert) architecture which allows them to fit into 80GB and 16GB GPU respectively, which enables them to be run with comparable low resources. Performance benchmarks read impressive too, so I can’t wait to get some time to experiment with them.

In their very insightful model card [1] OpenAI highlights that this is a model, not a system, card and that additional features, like system-level protections, typically provided by model serving providers are not part of the model. Developer and system integrators need to review use cases and decide if those need to be added on top. This is an important call out!

In my car analogy, I started to compare cars with GenAI applications some time ago already. In this analogy models become the engine. While most of the car buyers are caring little about the engine, it became commodity for many use cases of cars, buyers care for the overall feature set of the car, including safety and usability of the car. Staying in the analogy, we just got two very powerful and efficient new engine options which we can utilise under Apache 2.0 license. This is a great place to be!

I’m also excited about the announcement of the partnership between OpenAI and AWS [2] which allows builders to access both models in the established AWS managed services Amazon Bedrock and Amazon SageMaker from day 1 on. Specifically the integration into Amazon Bedrock allows builders to integrate the new models into their application, building performant & robust applications and agents, without getting lost in implementing all the utility parts which are required to turn an engine in a car. Amazon Bedrock Guardrails being just one example for a capability provided by Bedrock which I’m mentioning here because it is part of a direct response to the missing system-level controls highlighted in the OpenAI model card. The list goes on with mange’s infrastructure, knowledge bases, agent frameworks, …

With great power comes great responsibility. I’m impressed with the efforts OpenAI put into understanding and mitigating the risk of publishing open weight models. Again the model card [1] provided a good introduction to this and there is a research paper[3] published to dive deeper into this. This deserves another post.

So now I can’t wait to see what builder are going to build on tip of those models. Please share your thought and experiments!

And now go build!

[1] OpenAI model card: https://lnkd.in/eUytGTzX [2] AWS Announcement: https://lnkd.in/eAe7edP7 [3] OpenAI research paper: https://lnkd.in/eKwqy-5J

---