Mira Murati's Thinking Machines Lab Launches Inkling AI: Everything You Need to Know

The artificial intelligence landscape in 2026 is undergoing a major shift away from monolithic, centralized cloud systems toward highly customizable solutions. A key driver of this evolution is Thinking Machines Lab, the high-profile AI startup founded by former OpenAI Chief Technology Officer Mira Murati. The company has officially launched Inkling AI, a powerful new entry into the open ecosystem that changes how enterprises and developers approach model deployment.

The launch of the Inkling AI model represents a critical milestone in AI news. Instead of competing solely on raw benchmark supremacy against closed-source giants, Thinking Machines Lab has developed an open-weight AI model that organizations can download, self-host, and fundamentally modify to suit their precise operational workflows. For developers and business leaders who are increasingly concerned about the soaring costs, strict usage limits, and data privacy vulnerabilities of proprietary cloud APIs, Inkling AI offers a transparent and flexible alternative built on the permissive Apache 2.0 AI model framework.

What is Inkling AI?

Inkling AI is a state-of-the-art, natively multimodal AI foundation model engineered by Thinking Machines Lab. Released on July 15, 2026, the model is designed to process multiple types of inputs simultaneously—specifically text, high-resolution images, and continuous audio waveforms—while generating highly structured text, executable code, and system actions.

Thinking Machines Lab built this system from scratch with a clear and distinct philosophy: one-size-fits-all, centrally managed AI models are fundamentally limited when applied to diverse, specialized corporate tasks. Inkling is not intended to act as an unchangeable cloud oracle. Instead, it serves as a highly adaptable base layer meant to be extended, fine-tuned, and seamlessly integrated into complex internal networks, agentic tool-calling setups, and advanced automation pipelines.

The model is optimized to support developers who need absolute control over their underlying software infrastructure. By delivering full access to the network weights via Hugging Face and releasing it under a standard Apache 2.0 license, Thinking Machines Lab provides an enterprise-ready foundation that completely removes vendor lock-in.

Key Features

The engineering behind Inkling focuses on production utility rather than chasing isolated academic metrics. The table below highlights the key features that distinguish this model in the busy landscape of AI models 2026.

1. Multimodal Inputs

Capability: Natively ingests text (UTF-8), pixel-based images (40px to 4096px), and audio files (16kHz WAV up to 20 minutes). Benefit: Eliminates the need for separate vision or audio transcription models, reducing pipeline complexity.

2. Massive Context Window

Capability: Supports up to 1,048,576 tokens in full local execution. Available at 64K and 256K tiers via the Tinker API. Benefit: Allows developers to parse entire code repositories, hours of audio logs, or hundreds of pages of legal documents in a single query.

3. Controllable Thinking Effort

Capability: Uses a variable float parameter ranging from 0.00 (instant non-thinking response) to 0.99 (deep multi-step reasoning). Benefit: Gives teams the power to dynamically balance operational compute costs, response latency, and output accuracy depending on the task.

4. Open Weights & Apache 2.0

Capability: Full network parameters are completely open to download, modify, and host on private infrastructure. Benefit: Commercial use is permitted without royalties or restrictive usage limits, ensuring complete data privacy.

5. Tinker Platform Integration

Capability: Deep native compatibility with Thinking Machines Lab's "Tinker" fine-tuning and optimization suite. Benefit: Enables simple, efficient domain-specific fine-tuning and behavioral calibration using personal hardware or private cloud instances.

Technical Specifications: Mixture of Experts (MoE) Architecture 

To understand the sheer scale of the Inkling AI model, it helps to review its architectural blueprint. The model relies heavily on a highly efficient Mixture of Experts (MoE) routing topology to balance massive capacity with low operational overhead.

Core Architecture & Scale

• Base Model: 66-layer decoder-only autoregressive transformer with a sparse MoE backbone.
• Parameters: 975 Billion total (41 Billion active per individual token).
• Routing: Top-6 routing scheme across 256 total experts, plus 2 shared experts active on every token.
• Attention: Hybrid configuration blending localized and global attention layers.

Context & Modalities

• Maximum Context Window: 1,048,576 tokens.
• Input Modalities: Text, Images (RGB/pixel formats), Audio (16kHz WAV format).
• Output Modalities: Structured Text, Source Code, Tool Calls (UTF-8).

System Formats & Licensing

• Numerical Formats: Full native support for BF16, MXFP8, and highly efficient NVFP4 formats.
• License Type: Apache 2.0 (Permissive Commercial Open-Weight).

Demystifying the Specs

• Mixture of Experts (MoE): Instead of forcing every single word or image patch to activate all 975 billion parameters, the model uses an intelligent routing network. For any given token, the model activates just 41 billion parameters. This approach provides the extensive knowledge base of a near-trillion-parameter system while maintaining the rapid speed and lower compute footprint of a much smaller model.
• Active vs. Total Parameters: The 41 billion active parameters mean that running inference requires significantly less raw computing power than a traditional dense model of similar size. This architectural design directly helps lower infrastructure costs for enterprise deployments.
• Quantization Formats (NVFP4 and MXFP8): By supporting modern low-precision numerical formats, the model can compress its vast size down to run on more accessible hardware configurations.

Under the Hood: How the Inkling Multimodal Pipeline Works 

Inkling processes data through a unified multimodal pipeline, allowing it to seamlessly reason over text, vision, and audio without relying on disconnected external software pieces patched together.

The Multimodal Pipeline

When a user submits an input, the model breaks down the data into a single hidden layer space using dedicated encoders:

1. Text Inputs are processed through a standard high-efficiency UTF-8 text tokenizer.
2. Visual Inputs are broken down using a hierarchical patch encoder that converts the pixels into mathematical tokens while preserving spatial context.
3. Audio Waves are transformed into discrete token sequences via a specialized waveform encoder.

Once all inputs are projected into this shared mathematical hidden space, they flow into the main 66-layer decoder-only transformer network. Here, the MoE gating mechanism determines which specific experts are best suited to handle the data. The text output or code block is then generated iteratively.

Controllable Thinking Effort

One of Inkling's most powerful software innovations is its explicit Thinking effort level system prompt component. Users can pass a floating-point value to direct the model's inner processing style:

• Low Effort (none to 0.2): The model operates as a fast, direct autoregressive generator. This is ideal for straightforward text formatting, high-speed customer chat answers, or simple data conversions.
• High Effort (0.7 to 0.99): The model allocates significant internal tokens to reason, self-correct, cross-reference data points, and verify its logic before surfacing a final answer. This is designed for complex programming tasks, auditing advanced code bases, mathematical proofs, and deep multimodal analysis.

Open Weights vs. Open Source AI

There is often confusion in the industry regarding the terms "Open Weights," "Open Source," and "Closed Models." Understanding these distinctions is crucial for technical teams evaluating their deployment options.

Weight Availability

• Open Weights (e.g., Inkling AI): Fully Accessible — Anyone can download the complete trained parameters.
• Open Source (Strict OSI): Fully Accessible — Parameters are completely open alongside all source code.
• Closed Models (e.g., Cloud APIs): Completely Hidden — Weights are stored securely on the provider's private cloud.

Training Data & Code

• Open Weights (e.g., Inkling AI): Often Proprietary — The exact training data mixtures and dataset curation source scripts remain undisclosed.
• Open Source (Strict OSI): Fully Public — The complete training datasets, filtering pipelines, and source tracking are public.
• Closed Models (e.g., Cloud APIs): Completely Hidden — Data sourcing remains a corporate secret.

Deployment Control

• Open Weights (e.g., Inkling AI): Absolute — Can be run locally, inside private networks, or fully offline.
• Open Source (Strict OSI): Absolute — Can be deployed anywhere without oversight or restriction.
• Closed Models (e.g., Cloud APIs): Zero — Restricted to the provider's official APIs and cloud availability rules.

Customization Depth

• Open Weights (e.g., Inkling AI): High — Supports deep layer-wise fine-tuning and full structural weight adjustment.
• Open Source (Strict OSI): High — Full modifications can be made across any part of the architecture.
• Closed Models (e.g., Cloud APIs): Low — Limited to superficial prompt tuning or basic top-layer fine-tuning APIs.

The Value of the Apache 2.0 License

Thinking Machines Lab has distributed Inkling under the Apache 2.0 license. This highly permissive license grants developers extensive rights:

• You can freely use the model for commercial or private applications without paying royalties.
• You are permitted to modify the weights, alter the code, and package it into proprietary downstream software products.
• It includes a clear patent grant, protecting developers from unexpected intellectual property disputes.

Real-World Enterprise Use Cases for Inkling AI 

Because Inkling AI combines flexible multimodal AI capabilities with open architecture, it fits neatly into an array of practical business and technical applications.

1. Multi-Step Agentic Tool Calling

Inkling's architecture excels at tool use and structural reasoning. By turning up the thinking effort parameter, developers can construct robust autonomous agents that safely interact with external computer files, execute terminal commands, query corporate databases, and handle multi-step workflows without breaking down.

2. High-Context Retrieval-Augmented Generation (RAG)

With its expansive 1-million-token context window, Inkling solves a major problem in corporate document processing. Instead of cutting massive PDF manuals, dense legal contracts, or historical financial files into tiny, disjointed text snippets, enterprises can feed entire document bundles straight into the model. This vastly improves the accuracy of the generated answers and minimizes harmful hallucinations.

3. Native Multimodal Analysis

Businesses can use Inkling to audit mixed-media records automatically. For instance, a customer support platform can ingest an entire multi-turn technical interaction consisting of text logs, uploaded system screenshots, and recorded voice calls. Inkling can process all three modalities inside its shared hidden space to pinpoint exactly why an application failed.

4. Advanced Coding Assistants

Inkling demonstrates exceptional proficiency in software development tasks, tracking closely with top open alternatives on coding benchmarks while utilizing significantly fewer tokens. Because it is open-weight, engineering teams can host the model completely within their private local networks. This allows developers to use automated code completion and security auditing tools without exposing their private intellectual property to third-party public clouds.

Developer Benefits

For engineers building production software, the Inkling AI model offers several massive practical advantages over renting access to closed black-box APIs.

• Tailored Customization on Tinker: Through the Tinker platform, developers can easily run domain-specific post-training. You can fine-tune the model to speak in your brand's unique voice, strictly output specific JSON schemas, or master highly complex, proprietary programming syntaxes.
• Significantly Lower Infrastructure Costs: Running a sparse Mixture of Experts model means you only pay for the 41 billion parameters activated during the query. When paired with advanced local inference frameworks like SGLang, vLLM, or Unsloth, teams can achieve extreme token throughput at a fraction of the cost of premium closed cloud APIs.
• Flexible Quantization Options: Recognizing that not every lab has access to massive data center clusters, the weights are optimized for multiple precision levels. Developers can deploy quantized variants like NVFP4, which reduces the required memory overhead significantly, allowing large-scale workflows to run on more affordable hardware setups.

Current Limitations

While Inkling AI is a highly flexible, robust framework, technology journalists must evaluate its current limitations objectively to give teams a clear picture.

• Early Ecosystem Friction: As a new release from a young startup, Inkling does not yet possess the decades of community optimization, plug-and-play extensions, and deep third-party platform integrations enjoyed by older, established open model families.
• High Full-Precision Hardware Bar: If you want to deploy the raw, unquantized BF16 checkpoint locally, the hardware requirements are steep. Running it at full precision requires a GPU cluster providing at least 2 Terabytes of aggregated VRAM—such as 8 combined NVIDIA B300 GPUs or 16 NVIDIA H200 GPUs.
• Text-Only Output Constraints: Even though Inkling is exceptionally gifted at ingesting and understanding complex visual layouts and spoken audio files, its native output modality is strictly limited to text and code. It cannot natively generate synthetic speech audio or draw new images from scratch.

Inkling AI at a Glance

The following reference block provides an overview of the core facts surrounding the initial release of the model.

Inkling Model Specifications

Core Metadata & Licensing

• Model Developer: Thinking Machines Lab, Inc.
• Official Release Date: July 15, 2026
• Core Software License: Apache 2.0 (Highly permissive; allows for commercial use, modification, and private distribution).
• Weights Distribution: Open Weights available via Hugging Face (unlike fully closed proprietary models, you can download and run this locally or in a private cloud).

Architecture & Computational Footprint

• Model Type: Multimodal Sparse Mixture-of-Experts (MoE)
• Total Structural Parameters: 975 Billion
• Active Per-Token Parameters: 41 Billion

What this MoE design means in practice: While the model has a massive 975B parameter "brain" to pull specialized knowledge from, it only fires up 41B parameters to process any single token. This gives you the reasoning depth of a near-trillion-parameter model with the speed and lower hosting costs of a much smaller 41B model.

Capabilities & Modalities

• Context Window: Up to 1,048,576 tokens (1M tokens). Perfect for digesting entire codebases, massive legal documents, or hours of audio.
• Input Modalities: Text, Images, and Audio (natively multimodal on input).
• Output Modalities: Text and Code (UTF-8 format).

Best-Suited Operational Tasks

Because of its open-weight nature, massive context window, and native tool-use training, this model is built from the ground up for:

1. Agentic Workflows: Handling multi-step reasoning, executing code, and autonomously managing complex, branching tasks.
2. Retrieval-Augmented Generation (RAG): Searching, parsing, and synthesizing highly accurate answers from massive corporate databases (thanks to that 1M token context window).
3. Internal Tool Calling: Smooth integration with APIs, software databases, and internal developer tools.
4. Private Local Fine-Tuning: Ideal for enterprises that cannot upload sensitive data to third-party clouds. The Apache 2.0 license lets you adapt Inkling specifically to your proprietary data.

Frequently Asked Questions

What is Inkling AI?

Inkling AI is a 975-billion-parameter, open-weight multimodal foundation model created by Thinking Machines Lab. It is designed to run efficiently by utilizing a sparse Mixture of Experts architecture that only activates 41 billion parameters per token.

Who created Inkling?

The model was created by Thinking Machines Lab, Inc., an artificial intelligence research and development company led by former OpenAI Chief Technology Officer Mira Murati.

Is Inkling open source?

Inkling is technically classified as an open-weight model distributed under the highly permissive Apache 2.0 license. While its complete weights and parameters can be downloaded, modified, and used commercially for free, the exact raw training data mixtures and preparation scripts remain proprietary.

What does open weights mean?

Open weights means that the final trained parameters and mathematical configurations of the neural network are made completely public and downloadable. This allows developers to run the model locally, inspect its architecture, and fine-tune it on their own servers, unlike closed cloud APIs where the model is entirely hidden behind a web portal.

Can businesses use Inkling for commercial purposes?

Yes. Because Inkling is released under the standard Apache 2.0 license, enterprises can use, modify, host, and sub-license the model for commercial applications without paying any royalties or dealing with restrictive user volume caps.

What license does Inkling use?

Inkling is explicitly licensed under the Apache License, Version 2.0. This is a popular, developer-friendly open-software license that permits wide commercial utility and provides explicit patent protection.

Does Inkling support image and audio inputs?

Yes, Inkling is natively multimodal. It accepts standard pixel-based image files and WAV audio recordings up to 20 minutes long directly into its primary context window, alongside standard text prompts.

What makes Inkling different from traditional models?

Unlike closed, rigid cloud models, Inkling gives developers complete infrastructure ownership. Additionally, its built-in variable thinking effort parameter lets teams dynamically trade speed against deep reasoning capability on a query-by-query basis.

Conclusion

The debut of Inkling AI marks a definitive turning point for Open AI models in 2026. Rather than trapping users in an expensive, closed loop of proprietary cloud APIs, Thinking Machines Lab has delivered a massive, enterprise-ready framework that puts control squarely back into the hands of developers and businesses. By combining native multimodal AI capabilities and a massive 1-million-token context window with the efficiency of a Mixture of Experts design, Inkling provides a practical path forward for teams seeking local data privacy and custom domain expertise. As the ecosystem around Thinking Machines Lab matures, this foundational architecture is poised to reshape how modern enterprises configure their internal intelligence systems.