Mira Murati's New Company Releases 975B Open-Weight Model, Targeting the Top Tier of AI Competition

A Major Move from the Former OpenAI CTO

Thinking Machines Lab, founded by former OpenAI Chief Technology Officer Mira Murati, has released an open-weight large language model with 975 billion parameters, sparking intense discussion across technical communities including Hacker News. This move not only marks the startup's formal entry into model delivery, but also adds a heavyweight contender to an already fiercely competitive open-source LLM landscape.

Mira Murati joined OpenAI in 2018 and served as VP of Engineering before becoming CTO. She was one of the key driving forces behind GPT-3, GPT-4, the DALL·E series, and the commercialization of ChatGPT. With a background spanning engineering management and product development, she is widely credited as a central figure in translating OpenAI's research into commercial products. During the OpenAI board crisis in November 2023, she briefly served as interim CEO — further underscoring her central role within the company. In September 2024, she announced her departure from OpenAI and founded Thinking Machines Lab, which quickly secured major funding from top-tier venture capital firms (reportedly valued at over $2 billion), with several former OpenAI researchers and engineers joining her. This 975B-parameter model release marks the first opportunity for the outside world to glimpse this secretive company's technical direction.

What 975B Parameters Actually Means

A Scale That Competes at the Global Frontier

At 975 billion parameters, this large language model places itself squarely among the most capable models in existence today. For context, well-known frontier models such as GPT-4, the largest variants of the Llama series, and several models using MoE (Mixture of Experts) architecture generally fall in the hundreds of billions to trillions of parameters range. A near-trillion-parameter scale typically implies greater knowledge capacity and reasoning potential — but also brings substantial challenges in terms of deployment and inference costs.

MoE Architecture Explained: MoE (Mixture of Experts) is one of the core techniques that has enabled large language models to scale dramatically in recent years. The basic idea is to divide the model into multiple "expert" sub-networks, activating only a small subset of them for each input rather than running all parameters simultaneously. Google's Switch Transformer and GLaM were early representatives of large-scale MoE models, while more recent examples like Mixtral 8x7B — and GPT-4, which is widely rumored to use MoE — have further validated the architecture's engineering feasibility. For a model at the scale of 975B parameters, MoE is particularly important: if only roughly 1/8 of the parameters are activated per inference, the actual computational cost could be comparable to a dense model with tens of billions of parameters, significantly lowering the barrier to deployment.

The Strategic Logic Behind Open-Weight Releases

Notably, this release uses an open-weight model, not a fully open-source one — and the distinction matters. A fully open-source release typically means that model weights, training data, training code, evaluation pipelines, and more are all made publicly available, consistent with OSI's (Open Source Initiative) traditional definition of open-source software. An open-weight release, by contrast, only makes the trained model parameters publicly available for download, inference, and fine-tuning — without necessarily disclosing training data sources, full training frameworks, or other technical details. This is precisely why Meta's Llama series is considered by some researchers to be an "open model" rather than a truly open-source one. Developers can download and use the weights for inference, fine-tuning, and downstream development, but core assets such as training data and complete training code are not necessarily disclosed. This model has become the shared approach of leading players like Meta Llama and Mistral — leveraging the community ecosystem to expand reach while retaining key commercial moats.

For a startup, choosing the open-weight path is an ambitious move: it helps rapidly build a developer community, validate real-world model capabilities, and establish technical mindshare against closed-source giants.

Community Interest and Core Skepticism

The announcement garnered 95 upvotes and 18 comments on Hacker News, reflecting strong interest from the technical community. Discussion centered around three main areas:

Real-World Performance: Parameter count doesn't directly equal capability — training data quality, architectural design, and training strategies matter just as much. Developers are broadly eager to see how this model performs on standard benchmarks. These tests cover different dimensions: MMLU (Massive Multitask Language Understanding) spans 57 subject areas in multiple-choice format, testing broad knowledge; GSM8K consists of 8,500 math word problems specifically designed to evaluate chain-of-thought reasoning; and HumanEval contains 164 programming problems to measure code generation ability. It's also worth noting that as training data coverage grows ever broader, "benchmark contamination" — where test questions appear in training data — has become a significant concern affecting evaluation credibility, and the community remains vigilant on this front.

Deployment Barriers: A near-trillion-parameter model places extremely high demands on hardware. Inference may require multiple high-end GPUs or dedicated accelerator clusters, making it more suited for enterprises and research institutions than individual developers in the near term. Whether the model uses a MoE architecture to reduce active parameter counts at inference time has also become a key point of community discussion.

Commercial Sustainability: While open weights can generate strong ecosystem effects, training a model at this scale is enormously expensive. How Thinking Machines Lab plans to achieve commercial viability — through API services, enterprise customization, or other means — remains to be seen.

A New Variable in the Open-Source LLM Landscape

From a broader perspective, Thinking Machines Lab's entry further enriches the global competitive map of open-source large language models. Over the past two years, the open-source LLM ecosystem has undergone a structural shift from closed-source dominance to a parallel open-and-closed landscape. Meta's Llama 2 and Llama 3 series remain the most widely adopted open-weight models in the community, while French startup Mistral has earned broad recognition for its exceptional parameter efficiency. From China, Alibaba's Qwen series, DeepSeek^Ad's series, and 01.AI's Yi series constitute a significant and growing force in global competition — with DeepSeek-R1's reasoning performance in particular drawing substantial attention from the international community. Hugging Face, as the central hub for model hosting and distribution, provides the foundational infrastructure for the entire open-source ecosystem and now hosts over a million models.

The fact that a company founded by a core former OpenAI figure has chosen to go the open-weight route reflects, to some extent, the deep commitment that top AI talent has to the value of openness. For the broader industry, the availability of more high-quality open-weight models means that smaller companies and researchers can access near-frontier AI capabilities at lower cost, driving large-scale proliferation of application innovation. That said, a model's true value ultimately needs to be validated through thorough community testing.

Conclusion

Mira Murati and Thinking Machines Lab have made their formal debut with this 975B-parameter open-weight model — undoubtedly one of the most noteworthy developments in AI in recent memory. It carries the deep technical heritage of a former core OpenAI team, while also facing a multi-layered test of performance validation, deployment costs, and business model viability. As more benchmark results and community feedback emerge, we will be better positioned to make an accurate judgment about this model's true capabilities. Regardless of the outcome, a new player led by top industry talent and firmly committed to openness has injected fresh energy into the competitive landscape of large language models.

Key Takeaways