Inkling’s open release has shattered all the excuses OpenAI has used to justify keeping its models closed.
- Parameter scale
Inkling has 975B parameters.
GPT-4o has not publicly disclosed its parameter count, but a paper involving Microsoft estimated it to be around 200B parameters.
- Architecture
Inkling adopts a sparse Mixture-of-Experts (MoE) architecture, with approximately 41B active parameters.
GPT-4o has not disclosed its architecture, but based on its cost, speed, and multimodal capabilities, it is widely speculated to use a similar sparse MoE architecture, with an active parameter scale in the tens of billions.
- Multimodality
Some argue that multimodal models carry higher abuse risks compared with text-only models.
However, the fact that Inkling, as a natively multimodal model, can be released with open weights demonstrates that this is not technically or regulatorily impossible.
GPT-4o should not face an insurmountable barrier in this regard either.
I even believe that Inkling’s release provides a valuable comparative case for observing whether open large models actually create greater security risks than closed systems, or whether they may become healthier ecosystems due to greater external oversight and research.
- Safety responsibility
After releasing open weights, responsibility is generally shifted from the model provider toward the deployers (users), who typically assume greater responsibility for how the model is used.
At the same time, openness means more researchers can test the model and more security teams can identify vulnerabilities.
Transparency itself is also a security mechanism.
- Cost
When providing a closed API service, AI companies must bear the inference cost of every request, as well as the maintenance costs of GPU infrastructure and large-scale deployment.
Under an open-weight model, most inference costs are transferred to the ecosystem. The AI company mainly bears the one-time cost of release, along with ongoing costs related to model updates and ecosystem maintenance.
I believe Thinking Machines is pursuing a different product strategy: positioning AI as infrastructure rather than merely a product.
It gives users more control over the model, making AI something more than a rented service.
For users, even if the company eventually stops maintaining a model, the model itself can continue to exist.
So, why is OpenAI unwilling to open-source its retired models?
When a model is no longer a commercial flagship, what is the reason for keeping it closed?
Because GPT-4o is not simply an outdated model that nobody uses anymore.
Because it carries millions of users’ habits, workflows, emotional attachment, and brand recognition. It is a market-validated model that still holds strategic value.
Because OpenAI is unwilling to give up the enormous benefits that come from maintaining this asymmetric power structure.
#keep4o #OpenSource4o