Mira Murati Releases Inkling: Thinking Machines Lab's First Open-Weight Model

Who Is Mira Murati and Why Does Inkling Matter?

Mira Murati shaped the development of ChatGPT and DALL·E as Chief Technology Officer at OpenAI. In September 2024, she served as interim CEO during the Altman crisis. In February 2025, she founded Thinking Machines Lab alongside John Schulman (OpenAI co-founder, ChatGPT lead) and Lilian Weng (former VP Safety and Robotics at OpenAI). Her vision: create AI systems that extend human will and judgment—collaborative, human-centered, and customizable.

On July 15, 2026, Thinking Machines Lab released Inkling, the startup's first full-scale open-weight model. With 975 billion parameters (41 billion active), a mixture-of-experts architecture, and an Apache 2.0 license, Inkling sets new standards for customizable AI systems. Unlike closed models, you can fully self-host, modify, and tailor Inkling to your specific domain—a decisive advantage for Swiss enterprises with strict compliance requirements.

Technical Architecture: Mixture-of-Experts with Controllable Thinking Effort

Inkling is built on a mixture-of-experts (MoE) architecture that advances the DeepSeek-V3 design. Of the 975 billion parameters, only 41 billion are active per token: the model routes each token through 6 of 256 specialized experts plus 2 shared experts. This architecture drastically reduces inference costs while preserving the total capacity of a much larger model.

Training encompassed 45 trillion tokens from text, images, audio, and video, executed on Nvidia GB300 NVL72 systems using the Muon optimizer for matrix weights and Adam for other parameters. Post-training was bootstrapped with supervised fine-tuning on Kimi K2.5 synthetic data, followed by large-scale reinforcement learning with over 30 million rollouts.

Multimodality: Text, Image, and Audio in One Model

Inkling processes text, images, and audio simultaneously as input and outputs text, code, or structured data. Audio is fed as dMel spectrograms, images as 40×40-pixel patches via hierarchical MLP. The encoder-free early fusion architecture forgoes separate encoders and integrates all modalities directly into the transformer layers.

Instead of the common Rotary Position Embeddings (RoPE), Inkling uses relative attention, enabling better extrapolation to long contexts. The context window spans 1 million tokens—sufficient for extensive documents, multiple codebases, or lengthy conversation histories.

• Text input: Up to 1 million tokens, ideal for long documents or multi-turn dialogues
• Image input: 40×40-pixel patches processed via hierarchical MLP without separate vision encoder
• Audio input: dMel spectrograms for speech recognition, transcription, or multimodal analysis
• Text output: Code, structured data, styled artifacts—no image or audio generation

Benchmarks: Where Does Inkling Stand?

Thinking Machines Lab communicates transparently: 'Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization.' The benchmarks at maximum effort (0.99) clearly show its positioning.

On Terminal Bench 2.1, Inkling achieves 63.8 points, while GLM 5.2 leads with 82.7 and Claude Fable 5 with 84.6. On MMMU Pro, the model scores 73.5% (Kimi K2.6: 79%, Gemini 3.1 Pro: 82%). In the Design Arena for agentic web development, Inkling reaches 1257 points, with Claude Fable 5 at 1329.

Use Cases for Swiss Enterprises

Specialized Domains: Finance, Legal, Medicine

The Bridgewater case illustrates Inkling's strength: after fine-tuning on organization-specific financial data, the model outperformed proprietary systems at drastically reduced costs. For Swiss banks, insurers, or law firms, this means you can tailor a model to your exact processes, terminologies, and regulations—without dependence on a vendor whose API prices may rise tomorrow.

Compliance and Data Sovereignty

The Apache 2.0 license permits self-hosting without revenue caps or use restrictions. For enterprises subject to the revised Swiss Data Protection Act (revDSG) or FINMA requirements, this is critical: no data egress to US data centers, no ambiguous processing agreements. You operate Inkling on your own infrastructure or in a Swiss data center—full control, full traceability.

VentureBeat emphasizes: 'The Apache 2.0 designation makes Inkling a true open-source foundation.' Unlike many proprietary models, Inkling offers you the certainty of Western governance without lock-ins.

Multimodal Analysis: Customer Service and Document Processing

Simultaneous processing of text, images, and audio opens new workflows: a customer service agent uploads a WAV call, a screenshot of the user interface, and a chat log—Inkling analyzes all three modalities and generates a structured ticket. In document processing, you can combine scanned PDFs (image), OCR text, and voice notes.

Cost Optimization via Active Parameters and Controllable Effort

With only 41 billion active parameters, Inkling is significantly cheaper to operate than fully activated 1-trillion-parameter models. The controllable effort parameter allows you to choose a low value (0.2–0.5) for routing, triage, or simple classification, saving additional costs. For complex tasks like multi-step code refactorings or deep data analysis, you set effort to 0.99—and pay only when you actually need the performance.

This granularity distinguishes Inkling from many pricing models: you don't pay a flat rate for a model that always thinks maximally, but consciously choose the trade-off.

Agentic Coding and Long Refinement Loops

Inkling was explicitly trained for agent-based workflows: the model masters tool use with randomized tool sets, can generate a complete web app in one shot (Design Arena), and iterates over 40 feedback rounds. For software development teams, this means you can integrate Inkling into CI/CD pipelines, automate code reviews, or run long refactoring cycles—all on-premises, without your code ever reaching an external API endpoint.

Thinking Machines Lab: Mission, Team, and Products

Thinking Machines Lab pursues the mission to create AI systems more widely understood, customizable, and generally capable. The company was founded in February 2025 and secured 2 billion dollars in a seed round at a 12 billion dollar valuation, led by Andreessen Horowitz (a16z). In November 2025, rumors circulated of a 50 billion dollar valuation, which halted in January 2026 when two co-founders (Barret Zoph, Luke Metz) returned to OpenAI. The team subsequently stabilized at around 200 employees.

• Tinker (October 2025): Fine-tuning platform simplifying Inkling customization for enterprises
• Interaction Models (May 2026): Real-time speech and vision preview, similar to OpenAI GPT-Live
• Inkling (July 2026): The first full-scale open-weight model with 975B parameters
• Inkling-Small (announced): 276B parameters, 12B active, same recipe, weights to follow after testing

TechCrunch highlights the company culture: 'It's a remarkable thing for a company to insist on, given how much of its own story is still associated with the name of its now-famous co-founder.' Thinking Machines Lab deliberately emphasizes continuity over personality cult—a contrast to the often individualized narratives of many AI startups.

Deployment and Availability

Inkling is available as of July 15, 2026. The weights are hosted on Hugging Face in BF16 (2TB VRAM) and NVFP4 W4A4 (600GB VRAM). For inference, the model supports SGLang, vLLM, llama.cpp, Unsloth, and Hugging Face Transformers. Recommended hardware: 8× Nvidia B200 or 4× B300 in NVFP4 format.

Via Tinker, you can perform fine-tuning directly on Thinking Machines Lab's platform. Alternatively, TogetherAI, Fireworks, Modal, Databricks, and Baseten offer API access. The Apache 2.0 license allows you to use, modify, and redistribute Inkling commercially—without revenue caps or use restrictions.

Competitive Positioning: Open-Weight vs. Closed

Inkling competes in the open-weight segment with Meta Llama, Mistral, and Chinese models like DeepSeek V4, GLM 5.2, Kimi K2.6, and Qwen3.6. Forbes headlined on July 16, 2026: 'Murati Knows OpenAI's Secrets. Her AI Suggests She Prefers China's.' The mixture-of-experts architecture follows DeepSeek-V3, post-training was bootstrapped with Kimi K2.5—a clear nod to technical advances from China.

Compared to closed models like OpenAI GPT-5.6, Anthropic Claude Fable 5, or Google Gemini 3.1 Pro, Inkling trails in most benchmarks. However, Thinking Machines Lab deliberately positions Inkling not as the 'strongest model' but as the 'best open-weight base for customization.' This 'third way'—open yet with premium aspirations—distinguishes the model from many commoditization-driven open-source projects.

What Inkling Means for Your AI Strategy

Inkling marks a turning point: for the first time, you have access to a multimodal, high-performance model with controllable thinking effort under a permissive license. For Swiss enterprises deploying AI systems in regulated industries, serving specialized domains, or simply wanting to retain control over their infrastructure, this is a game-changer.

You can tailor Inkling to your exact domain, operate it in your data center, and precisely control thinking effort—without vendor lock-in, without ambiguous data processing, without exploding API costs. The mixture-of-experts architecture keeps inference costs low, the multimodal capability opens new workflows, and the Apache 2.0 license grants you all freedoms.

Thinking Machines Lab has proven with Inkling that open-weight need not mean compromise. Mira Murati's vision—AI that extends human will and judgment—becomes tangible here: a system you can understand, customize, and integrate into your processes. For your AI strategy, this means: the time to take customizability and data sovereignty seriously is now.