Key Points:
- AI startup Thinking Machines, founded by former OpenAI CTO Mira Murati, launched its first proprietary model called Inkling.
- Inkling is a 975-billion-parameter open-weight model utilizing a Mixture of Experts (MoE) architecture with 41 billion active parameters.
- The multimodal model reasons across text, images, and audio, trained on 45 trillion tokens with up to a 1 million token context window.
- It provides a powerful Western alternative to dominant Chinese open models, filling a critical gap in the US open-weights ecosystem.
A major transition is underway in the artificial intelligence industry as Silicon Valley startup Thinking Machines launches its first proprietary AI model. The newly unveiled open-weight model, named Inkling, directly challenges the dominant subscription-only APIs offered by market incumbents. By distributing the underlying model weights openly, the company is betting that enterprise clients prefer to download, host, and customize their own artificial intelligence systems rather than relying on a small handful of closed-source frontier labs.
The open-weight model features a highly sophisticated Mixture of Experts (MoE) architecture designed to balance computational efficiency with high-performance reasoning. The system contains a massive 975 billion total parameters, with 41 billion parameters actively running during any single computational query. Pre-trained on a colossal dataset of 45 trillion tokens containing text, images, audio, and video, the model reasons natively across multiple modalities. This multimodal design allows the software to process, analyze, and generate highly complex outputs spanning written documents, voice files, and visual graphics simultaneously.
To support extensive data-processing tasks, the new model supports an exceptionally large context window of up to 1 million tokens. Developers can access the model with initial context length options of 64,000 and 256,000 tokens through the company’s Tinker customization platform. Tinker, which originally launched in October 2025 as a dedicated API for fine-tuning open-source software models, has immediately integrated the new model card. This integration allows companies to immediately run low-rank adaptation fine-tuning on their own private data, securing complete control over their proprietary intellectual property.
The launch of the 975-billion-parameter model addresses a critical structural gap in the Western technology ecosystem. Over the past several months, the supply of high-performance open-weight models from American developers had looked increasingly thin. This scarcity grew more pronounced after social media giant Meta shifted its focus away from open standards following a lackluster reception for its Llama 4 family, leaving many Western enterprises with no choice but to adopt powerful open-weight models coming out of Chinese AI labs. The arrival of the new model provides a highly competitive, domestically developed alternative for organizations that require self-hosted software.
While the newly released system does not claim to beat the smartest, highly expensive closed-source models across every single benchmark, it delivers exceptional, competitive results across a wide range of practical applications. The software performs particularly well on agentic workloads, where digital assistants must independently plan, execute, and troubleshoot multi-step actions across various databases and APIs. This technical strength makes it a highly attractive, cost-effective substitute for companies looking to deploy complex, autonomous software agents on their own internal infrastructure.
The release of this model aligns directly with a human-centric philosophy of technology development. Artificial intelligence should extend human will and judgment rather than simply freezing a snapshot of organizational knowledge to replace the employees who generated it. By prioritizing customizable, open-weight models that companies can adapt to the full spectrum of human expertise, the startup is positioning itself as a strategic, philosophical counterweight to the highly centralized, closed-source models dominating the market.
The rapid rise of the San Francisco-based startup has made it one of the most closely watched corporate entities in Silicon Valley. Founded in February 2025 by former OpenAI Chief Technology Officer Mira Murati, the company completed a historic, record-shattering $2 billion seed funding round in July of that year. The massive financing, which valued the pre-product startup at an unprecedented $12 billion post-money, was led by prominent venture capital firm Andreessen Horowitz. Other notable strategic backers included major chipmakers Nvidia and AMD, networking giant Cisco, and market-making firm Jane Street.
The founding team consists of elite researchers and engineers poached directly from competitors like OpenAI, Meta, and Mistral AI. Co-founder John Schulman, a highly respected computer scientist who co-founded OpenAI, serves as the startup’s Chief Scientist. Although the company experienced some high-profile talent departures, including the return of several research directors to OpenAI and a failed $1 billion takeover attempt by Meta that the board decisively rebuffed, the core engineering team of approximately 100 professionals has remained highly focused on delivering functional software.
The availability of a highly capable, self-hosted alternative fundamentally reshapes the dynamics of enterprise technology acquisitions. For a long time, major cloud providers held a near-monopoly on high-end intelligence, allowing them to dictate pricing and data policies to corporate clients. The arrival of a massive, 975-billion-parameter open-weight model that is good enough to handle enterprise-level tasks gives corporate buying teams significant leverage. Companies can now credibly threaten to self-host their own models on private server infrastructure, forcing major API providers to lower their pricing.
Ultimately, the commercial debut of Inkling represents a defining moment for the open-weight AI ecosystem. By delivering a massive, multimodal model that companies can customize and host natively, the startup has delivered a powerful alternative to the centralized, subscription-only software networks. As the competition between open-source models and closed-source APIs continues to accelerate, the success of this 975-billion-parameter model will demonstrate whether the future of artificial intelligence belongs to a handful of cloud-based giants or a decentralized, highly customizable open ecosystem.
