I cannot believe I am saying this, but OpenAI actually went and open-sourced a model. After years of keeping everything behind API walls, they dropped GPT-OSS-120B on us in August with Apache 2.0 license. Pretty wild when you think about it.
The thing is, this is not some toy model they threw over the fence either. It benchmarks close to o4-mini, which is their own reasoning model. Input costs around 0.8 yuan per million tokens — that is dirt cheap compared to what you would pay for equivalent capability through their API. You can run it locally, fine-tune it, throw it in products. Commercial use included.
But here is where it gets interesting. Running a 120B parameter model locally is not something most people can just do on their laptop. You would need serious GPU hardware, probably 8x H100s or something in that neighborhood. The inference performance is close to o4-mini, sure, but you are looking at a significant infra investment to actually host it yourself.
I have not personally benchmarked it against running the same workload through OpenAI API yet. My gut says for most teams, the API route might still make more sense financially when you factor in the operational overhead of self-hosting. But the option is there now, and that is a meaningful shift for the ecosystem.
What is worth noting is the timing. They announced this alongside an $8 billion financing round at $300 billion valuation. Mix of PR play and genuine strategic move — they are betting that opening up the weights creates an ecosystem lock-in even if someone else matches their closed products eventually.
The open-source LLM space is getting crowded though. Meta LLaMA has been there, Mistral has been pushing releases, and now this. For practitioners, it is good trouble we are having here — more choices, more competitive pricing, more ability to self-host. I just hope they keep iterating on it rather than dropping it and walking away like some companies do.
This is genuinely worth keeping an eye on if you have been waiting for a legitimate OpenAI-backed open weights model to experiment with.
