Thinking Machines Lab’s New Inkling Model and the Tinker Ecosystem.

Mira Murati’s 'Thinking Machines Lab' Debuts Inkling: A Powerhouse 975B Open-Weight Multimodal AI Model

Thinking Machines Lab the highly anticipated AI startup founded by former OpenAI Chief Technology Officer Mira Murati has officially stepped out of stealth mode with the release of its inaugural flagship model, Inkling. Trained entirely from scratch, Inkling is a native multimodal powerhouse built on a highly customizable open-weight architecture, signaling the startup's commitment to democratic, developer-first AI infrastructure.

Architecturally, Inkling boasts a massive 975 billion total parameters, utilizing a sparse routing mechanism that maintains an efficient 41 billion active parameter threshold during inference. Built to handle complex, long-form data ingestion, the model supports an expansive context window of up to 1 million tokens with native, synchronous processing capabilities across text, vision, audio, and video inputs.

The core selling point of Inkling is its unprecedented layer of customizability. Developers can dynamically recalibrate the model's weights to optimize it for highly specific downstream tasks. This granular configuration is designed to interface seamlessly with Tinker, an open-weight fine-tuning orchestration platform previously developed and open-sourced by Thinking Machines Lab.

In raw, zero-shot benchmarking suites, Inkling delivers a balanced, mid-tier performance matrix, lacking the immediate chart-topping scores found in rigid proprietary models or heavily distilled open alternatives. However, Thinking Machines Lab clarified that this is a deliberate trade-off; by maximizing downstream weight flexibility, the architecture achieves elite infrastructure efficiency, allowing developers to attain highly optimized, cost-effective performance post-fine-tuning.

Alongside the primary engine, the startup provided a technical preview of Inkling-Small. Featuring 276 billion total parameters (12 billion active parameters), the distilled variant closely mirrors the core capabilities of its larger sibling at a fraction of the compute cost, making it an ideal candidate for targeted tasks like high-speed code generation. Both models are available for immediate deployment via Tinker and Hugging Face.

The Inkling Model Architecture Blueprint

• The Pedigree: The debut model from Thinking Machines Lab, founded by legendary ex-OpenAI CTO Mira Murati.

• The Framework: Fully open-weight architecture, trained entirely from scratch.

• Parameter Scaling:

  • Inkling (Main): 975B total / 41B active parameters.

  • Inkling-Small: 276B total / 12B active parameters (optimized for low-latency coding tasks).

• Context & Modality: 1-million token context window with native processing for text, images, audio, and video.

• Customization Engine: Fully compatible with Tinker, the lab's proprietary open-weight tuning interface.

• Availability: Live right now for open deployment on Hugging Face and the Tinker platform.

Behind the scenes (Founder Moat): As the former CTO of OpenAI, Murati played a key role in bringing ChatGPT, GPT-4, and cutting-edge models like o1 to the world stage. Her choice to launch her new company with an open-weight model instead of a closed, monthly subscription-based system (SaaS model) is a powerful statement. She envisions the future of generative AI not being a monopoly on technology in an ivory tower, but rather providing "back-end keys" (weights) to developers and large organizations so they can customize, expand, and run them on private clouds with secure data – a completely different selling point from her previous approach.

Sparse Parameter Activation: The figure of 975 billion parameters indicates the model's immense cognitive capacity, capable of storing deep knowledge substrates and linguistic dimensions. However, the fact that it only actively activates 41 billion parameters reflects the use of a Mixture of Experts (MoE) architecture, where the model's brain is divided into smaller departments as user inputs are received. The system selectively wakes up only the departments that are specialized in a particular area, resulting in very low compute costs. This solves the classic problem of large models that are often slow and resource-intensive.

Inkling's raw test results are average, not spectacular, but in the open-source AI world, developers don't need models that are "all-around excellent" from the start. They need models with highly flexible structures, free from bias from excessive tweaking. Thinking Machines Lab's strategy of pairing Inkling with the Tinker platform demonstrates their goal of creating the best possible foundational layer. This allows organizations to fine-tune the model, adding specialized data such as medical or legal information. Ultimately, this fine-tuning will make the specialized model more versatile than a general-purpose closed-system model.

 

 

Australia Mandates 3 Hours of Free Daily Electricity as Rooftop Solar Overwhelms the Grid. 

 

Source: Thinking Machines Lab 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Mira Murati’s 'Thinking Machines Lab' Debuts Inkling: A Powerhouse 975B Open-Weight Multimodal AI Model

Thinking Machines Lab the highly anticipated AI startup founded by former OpenAI Chief Technology Officer Mira Murati has officially stepped out of stealth mode with the release of its inaugural flagship model, Inkling. Trained entirely from scratch, Inkling is a native multimodal powerhouse built on a highly customizable open-weight architecture, signaling the startup's commitment to democratic, developer-first AI infrastructure.

Architecturally, Inkling boasts a massive 975 billion total parameters, utilizing a sparse routing mechanism that maintains an efficient 41 billion active parameter threshold during inference. Built to handle complex, long-form data ingestion, the model supports an expansive context window of up to 1 million tokens with native, synchronous processing capabilities across text, vision, audio, and video inputs.

The core selling point of Inkling is its unprecedented layer of customizability. Developers can dynamically recalibrate the model's weights to optimize it for highly specific downstream tasks. This granular configuration is designed to interface seamlessly with Tinker, an open-weight fine-tuning orchestration platform previously developed and open-sourced by Thinking Machines Lab.

In raw, zero-shot benchmarking suites, Inkling delivers a balanced, mid-tier performance matrix, lacking the immediate chart-topping scores found in rigid proprietary models or heavily distilled open alternatives. However, Thinking Machines Lab clarified that this is a deliberate trade-off; by maximizing downstream weight flexibility, the architecture achieves elite infrastructure efficiency, allowing developers to attain highly optimized, cost-effective performance post-fine-tuning.

Alongside the primary engine, the startup provided a technical preview of Inkling-Small. Featuring 276 billion total parameters (12 billion active parameters), the distilled variant closely mirrors the core capabilities of its larger sibling at a fraction of the compute cost, making it an ideal candidate for targeted tasks like high-speed code generation. Both models are available for immediate deployment via Tinker and Hugging Face.

The Inkling Model Architecture Blueprint

• The Pedigree: The debut model from Thinking Machines Lab, founded by legendary ex-OpenAI CTO Mira Murati.

• The Framework: Fully open-weight architecture, trained entirely from scratch.

• Parameter Scaling:

  • Inkling (Main): 975B total / 41B active parameters.

  • Inkling-Small: 276B total / 12B active parameters (optimized for low-latency coding tasks).

• Context & Modality: 1-million token context window with native processing for text, images, audio, and video.

• Customization Engine: Fully compatible with Tinker, the lab's proprietary open-weight tuning interface.

• Availability: Live right now for open deployment on Hugging Face and the Tinker platform.

Behind the scenes (Founder Moat): As the former CTO of OpenAI, Murati played a key role in bringing ChatGPT, GPT-4, and cutting-edge models like o1 to the world stage. Her choice to launch her new company with an open-weight model instead of a closed, monthly subscription-based system (SaaS model) is a powerful statement. She envisions the future of generative AI not being a monopoly on technology in an ivory tower, but rather providing "back-end keys" (weights) to developers and large organizations so they can customize, expand, and run them on private clouds with secure data – a completely different selling point from her previous approach.

Sparse Parameter Activation: The figure of 975 billion parameters indicates the model's immense cognitive capacity, capable of storing deep knowledge substrates and linguistic dimensions. However, the fact that it only actively activates 41 billion parameters reflects the use of a Mixture of Experts (MoE) architecture, where the model's brain is divided into smaller departments as user inputs are received. The system selectively wakes up only the departments that are specialized in a particular area, resulting in very low compute costs. This solves the classic problem of large models that are often slow and resource-intensive.

Inkling's raw test results are average, not spectacular, but in the open-source AI world, developers don't need models that are "all-around excellent" from the start. They need models with highly flexible structures, free from bias from excessive tweaking. Thinking Machines Lab's strategy of pairing Inkling with the Tinker platform demonstrates their goal of creating the best possible foundational layer. This allows organizations to fine-tune the model, adding specialized data such as medical or legal information. Ultimately, this fine-tuning will make the specialized model more versatile than a general-purpose closed-system model.

 

 

Australia Mandates 3 Hours of Free Daily Electricity as Rooftop Solar Overwhelms the Grid. 

 

Source: Thinking Machines Lab