Thinking Machines Launches Inkling AI Model with Open Weights

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first in-house AI model today.

The new AI system, named Inkling, uses an open-weight design that allows public researchers and companies to download and modify the code directly.

Credit: Thinking Machines

Inkling model features 975 billion total parameters and utilizes a Mixture-of-Experts architecture that activates 41 billion parameters for any individual task. It processes multimodal data including text, images, audio, and video, though it currently limits its final outputs to text and code.

Murati announced the product launch on the social media platform X, confirming that users can access the system immediately.

"Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today."

Tinker serves as the startup's platform where enterprise clients can customize large language models for specific business tasks. Thinking Machines is currently providing access to Inkling without charging any fees.

Model Specifications and Performance Limits

The company openly states that Inkling does not outperform every competitor on standard leaderboards. Company materials describe the release as not the strongest model available today among closed or open systems. Instead, the developers focused on customizability and lowering operational costs for companies that want to host models on their own private servers.

• Total Parameters: 975 billion
• Active Parameters: 41 billion
• Context Window: Up to 1 million tokens
• License: Apache 2.0 open source license

Benchmark tests provided by the company indicate that Inkling matches Nvidia’s Nemotron 3 Ultra on coding tests and uses a third of the tokens. The model scores behind closed proprietary flagship systems like OpenAI’s GPT-5.6 and Anthropic’s Claude Fable 5.

Technical Ties to Chinese Competitors

The technical development of Inkling involved drawing from systems developed by Chinese technology competitors. Thinking Machines built the foundation architecture based on China's DeepSeek-V3 system. The company also used post-training data generated by Moonshot AI’s Kimi K2.5 model, a development method called distillation.

The firm stated that its next generation of models will train entirely on internal data rather than using outputs from external models. The company trained the model entirely on Nvidia GB300 systems, following a hardware infrastructure agreement established in March.

Thinking Machines also released a preview version of a smaller model named Inkling-Small. This secondary model operates with 12 billion active parameters to reduce data latency and lower processing expenses.

Corporate Structure and Funding

Thinking Machines formed in February 2025 and secured two billion dollars in seed funding last year at a twelve-billion-dollar valuation. Investors included venture capital firm Andreessen Horowitz, chipmakers Nvidia and AMD, and the hedge fund Jane Street. The government of Albania also contributed a ten-million-dollar direct investment into the startup.

The San Francisco startup currently employs roughly 200 workers following the departure of two co-founders earlier this year. Business records show that Murati retains a majority voting share that grants her primary control over board decisions.