Thinking Machines publishes a 975-billion-parameter "open weights" model

Mira Murati, former CTO of OpenAI, releases a frontier model with open weights from Thinking Machines - something her former employer has been promising for years without delivering.

The context

Since OpenAI definitively switched to "closed source" with GPT-4, the "open weights" scene (models whose weights are downloadable, even if the license is not strictly free) has been dominated by the Chinese: Qwen (Alibaba), DeepSeek, Kimi. American players offered either closed models (OpenAI, Anthropic) or "small" open models (Meta with Llama, a few university initiatives).

Thinking Machines, the company founded by Mira Murati (the former CTO of OpenAI) after her dramatic departure, has just released its first open weights model. And they aimed high.

What's in the box

According to The Register, the model has 975 billion parameters. To put this into perspective, it's of the same order of magnitude as what rumors attributed to GPT-4 (around 1.8T with Mixture of Experts). The weights are downloadable, which allows:

• to run it on-premise (provided you have the necessary GPU farm),
• to fine-tune it on a domain (law, medical, code),
• to audit its behavior without depending on a third-party API.

The stated idea is to offer an American alternative to Chinese open models, in line with the position that Murati already defended at OpenAI—a position to which Sam Altman put an end when OpenAI became, in practice, a SaaS provider.

What this changes for a dev

Concretely, a model with 975B parameters, you don't run it on your laptop. We're talking about a high-end multi-GPU cluster, or an inference provider that will host it. What changes is:

• Reproducibility: the weights are frozen, you can version your AI as you version your code.
• Sovereignty: no longer need to leak sensitive data to OpenAI or Anthropic's API.
• Research: universities and independent labs can finally study a model of this size in its entirety.

Note: "open weights" ≠ "open source" in the OSI sense. The training data and the training code remain mostly closed. For truly free in the historical sense, you'll have to continue looking at projects like OLMo (AI2) or BLOOM.

Key takeaways

Mira Murati publicly delivers what Altman has been promising for two years without doing it. It's the strongest political signal of the year for the Western open AI scene—and a reminder that when a leader leaves with cash, they sometimes do what they couldn't do from the inside.