Mira Murati's Thinking Machines Lab Releases Inkling, a 975B Open-Weight Multimodal Model

Mira Murati's Thinking Machines Lab released Inkling, a 975-billion-parameter open-weight multimodal model that processes text, images, audio, and video. The model, trained on 45 trillion tokens, is available on Hugging Face and through API partners, with the company generating revenue via its Tinker customization platform rather than the model itself. Thinking Machines raised $2 billion at a $12 billion valuation and positions Inkling as a base for enterprise fine-tuning rather than a top-performing general-purpose model.

Mira Murati's Thinking Machines Lab Releases Inkling, a 975B Open-Weight Multimodal Model - Inkling is a 975B-parameter mixture-of-experts model with 41B active parameters, trained on 45 trillion tokens spanning text, images, audio, and video 1 https://thinkingmachines.ai/news/introducing-inkling/ - The model is open-weight and available on Hugging Face, with API access through Together AI, Fireworks, Modal, Databricks, and Baseten 1 https://thinkingmachines.ai/news/introducing-inkling/ - Thinking Machines does not monetize Inkling directly — revenue comes through Tinker, its model-customization platform, with clients including Bridgewater Associates 3 https://fortune.com/2026/07/15/what-is-mira-murati-thinking-machines-first-ai-model-inkling/ - The company raised $2 billion at a $12 billion valuation in July 2025, with investors including Nvidia, AMD, Andreessen Horowitz, and Jane Street 4 https://en.wikipedia.org/wiki/Thinking Machines Lab - Thinking Machines acknowledged Inkling is 'not the strongest model available today, closed or open,' positioning it instead as a base for enterprise fine-tuning 1 https://thinkingmachines.ai/news/introducing-inkling/ Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, on July 15 released Inkling — a 975-billion-parameter open-weight multimodal model that natively processes text, images, audio, and video. The model uses a mixture-of-experts architecture that activates only 41 billion parameters per query, and was trained on 45 trillion tokens across all four modalities 1 . The release marks the first public model from Thinking Machines, which has operated largely out of public view since Murati founded it in February 2025. The San Francisco-based company, structured as a public benefit corporation, raised $2 billion at a $12 billion valuation in a seed round led by Andreessen Horowitz, with participation from Nvidia, AMD, Cisco, Jane Street, and the government of Albania 4 . Inkling's weights are available for download on Hugging Face and accessible through API partners including Together AI, Fireworks, Modal, Databricks, and Baseten. The company is not charging for the model itself — instead, it generates revenue through Tinker, a platform that lets organizations fine-tune and customize AI models for their own use cases 1 https://thinkingmachines.ai/news/introducing-inkling/ 3 . The Model Inkling's architecture features 256 routed experts per mixture-of-experts layer plus two shared experts, with six routed experts active per token. It supports a context window of up to one million tokens and uses interleaved sliding-window and global attention layers at a 5:1 ratio 1 . On benchmarks, the model scores 97.1% on AIME 2026 and 87.2% on GPQA Diamond for reasoning tasks. In agentic coding, it achieves 77.6% on SWE-Bench Verified. For multimodal tasks, it reaches 73.5% on MMMU Pro and 91.4% on VoiceBench 1 . A smaller variant, Inkling-Small, is also available in preview with 276 billion total parameters and 12 billion active parameters. Thinking Machines says the smaller model offers comparable reasoning and coding performance but lower factuality than the full version 1 . The model includes a controllable 'thinking effort' dial that lets users trade accuracy for speed. At reduced effort levels, Thinking Machines claims Inkling matches Nemotron 3 Ultra performance at roughly one-third the token consumption on Terminal Bench 2.1 1 . Strategic Positioning Thinking Machines is making a deliberate bet against the prevailing approach of selling monolithic, general-purpose AI models. The company acknowledged in its release blog post that Inkling is 'not the strongest overall model available today, open or closed,' framing it instead as 'a good open-weights base for customization' 1 . The business model centers on Tinker, the company's fine-tuning platform launched in October 2025. Clients including hedge fund Bridgewater Associates use Tinker to adapt models for domain-specific tasks such as financial analysis. Tinker is currently offering a 50% limited-time discount, with 64K and 256K context options available 1 https://thinkingmachines.ai/news/introducing-inkling/ 3 . The open-weight approach positions Thinking Machines in contrast to OpenAI, which has kept its most capable models proprietary, and to Meta, which has shifted toward more restrictive licensing for its Llama models. The release comes as Chinese AI labs have gained ground in the open-source space, raising concerns among U.S. policymakers about competitiveness 3 . The Team and Backers Murati's team draws heavily from OpenAI's former leadership. Chief Scientist John Schulman is an OpenAI co-founder, while other hires include Barret Zoph, formerly OpenAI's vice president of research, and Lilian Weng, formerly OpenAI's vice president. Advisers include Bob McGrew and Alec Radford, also from OpenAI 4 . The company has grown from roughly 30 researchers and engineers at launch to around 100 to 170 employees, though it has faced some attrition. Several employees have departed for competitors including Meta and OpenAI 3 https://fortune.com/2026/07/15/what-is-mira-murati-thinking-machines-first-ai-model-inkling/ 4 . In March 2026, Thinking Machines struck a strategic partnership with Nvidia involving one gigawatt of computing capacity. The model was trained on Nvidia's GB300 NVL72 systems 1 https://thinkingmachines.ai/news/introducing-inkling/ 4 . An attempt to raise additional funding at a valuation of $50 billion to $60 billion in late 2025 did not close . 5 https://finance.yahoo.com/news/mira-muratis-thinking-machines-seeks-212328387.html What's Next The Inkling release serves as Thinking Machines' first major proof point after 18 months of development. The company's near-term focus is on expanding the Tinker platform's enterprise customer base and demonstrating that customizable open models can compete with the closed offerings from OpenAI, Anthropic, and Google 2 . Inkling is available immediately for inference through its API partners and for download via Hugging Face, including an NVFP4 checkpoint optimized for Nvidia Blackwell hardware. Compatible inference frameworks include SGLang, vLLM, llama.cpp, and TokenSpeed 1 . Companies mentioned Further sources 1 Thinking Machines Lab, 'Inkling: Our open-weights model,' official blog post, J… ↗ https://thinkingmachines.ai/news/introducing-inkling/ 2 TechCrunch, 'Thinking Machines amps up its bet against one-size-fits-all AI wit… ↗ https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/ 3 Fortune, 'Murati's Thinking Machines releases first AI model for broad use,' Ju… ↗ https://fortune.com/2026/07/15/what-is-mira-murati-thinking-machines-first-ai-model-inkling/ 4 Wikipedia, 'Thinking Machines Lab' ↗ https://en.wikipedia.org/wiki/Thinking Machines Lab 5 Yahoo Finance / Bloomberg, 'Mira Murati's Thinking Machines seeks $50 billion v… ↗ https://finance.yahoo.com/news/mira-muratis-thinking-machines-seeks-212328387.html The stories that matter, in one email. Free — unsubscribe anytime.