Thinking Machines Launches Inkling: 975B Open-Weight Model That Acknowledges Its Limits

On July 15, 2026, Thinking Machines Lab — the startup founded by Mira Murati, former CTO of OpenAI — launched Inkling, its first artificial intelligence model. It is a Mixture-of-Experts (MoE) system with 975 billion total parameters, though only about 41 billion are active per task. It was trained on 45 trillion multimodal tokens (text, image, audio, and video) on Nvidia GB300 NVL72 infrastructure, according to Connie Loizos at TechCrunch. The unusual part: the company itself acknowledges that Inkling is not the most powerful model on the market, open or closed. Instead of promising absolute performance, it highlights its ability to signal uncertainty (saying 'I don't know' instead of hallucinating), an adjustable 'thinking effort' dial, and efficiency that in code benchmarks consumes one-third the tokens of Nvidia Nemotron 3 Ultra for equivalent performance. Inkling is open-weight: the weights are available for download and modification. Thinking Machines' real business is Tinker, its customization and hosting platform. This launch breaks with the industry's tradition of hype: the company explicitly stated that 'Inkling is not the most powerful model available today, closed or open,' an unprecedented move in a sector where the norm is to promise maximum performance.

Background and Historical Context

Thinking Machines Lab was founded in 2025 by Mira Murati, who was CTO of OpenAI until 2024. Murati led the development of GPT-4 and DALL-E 3, but left the company after strategic differences over the direction of open versus closed AI. Her new lab raised over $1 billion in its first round, with investors including Andreessen Horowitz and Nvidia. The name 'Thinking Machines' evokes the mythical supercomputing company of the 1980s and 1990s, which pioneered parallel computing before going bankrupt. This historical nod underscores Murati's ambition to redefine AI through transparency and efficiency, in contrast to the closed approach of OpenAI and Google DeepMind. The launch of Inkling comes at a time when the AI industry is experiencing a 'hallucination crisis' and runaway inference costs. Closed models like GPT-5 (released in 2025) and Gemini 3 (2026) have prioritized raw performance, but with inference costs that can exceed $10 per million tokens for complex tasks. In this context, Inkling emerges as an alternative that prioritizes honesty and efficiency.

Why It Matters

Thinking Machines' move represents a strategic shift in the industry. While OpenAI, Google, and Anthropic compete for the most powerful closed model, Murati bets on transparency and efficiency. It's a sign that the AI market is maturing: companies no longer just want the best performance, but control, customization, and predictable costs. Moreover, uncertainty flagging addresses one of the most critical problems in AI today: hallucinations. If a model admits it doesn't know, users can trust its answers more. This could mark a before and after in enterprise adoption of open models. Token efficiency is another differentiating factor. In an environment where inference costs are a bottleneck, reducing consumption to one-third without losing performance is a huge competitive advantage. According to TechCrunch data, on the HumanEval benchmark for code generation, Inkling scored 82.3% versus 84.1% for Nemotron 3 Ultra, but consumed only 34% of the tokens. This implies a 66% cost savings on code tasks, which can be decisive for startups and companies with large inference volumes. In comparison, models like Llama 4 (Meta) consume approximately 80% of Nemotron's tokens for similar performance, positioning Inkling as an efficiency leader.

Market Consequences

Inkling pressures closed giants to justify their prices. If an open-weight model performs nearly as well on key tasks like code, companies might migrate to open solutions. It also reinforces the trend toward specialized models: instead of a one-size-fits-all model, the idea of base models that each company fine-tunes to its domain is gaining ground. For Nvidia, the alliance with Thinking Machines — which includes a gigawatt of Vera Rubin capacity — consolidates its dominance in AI infrastructure. For competing startups like Mistral or Meta, Inkling is a direct rival in the open-weight space. Mistral recently launched Mistral Large 3 with 500 billion parameters, but its efficiency is lower: it consumes twice the tokens of Inkling on reasoning tasks. Meta, meanwhile, has announced Llama 5 for late 2026 but has not yet detailed its architecture. Thinking Machines' success will depend on Tinker adoption. The customization platform is where the company expects to generate revenue, competing with services like Hugging Face or AWS SageMaker. Tinker offers domain-specific fine-tuning (legal, medical, finance) with prices starting at $0.05 per GPU hour, significantly cheaper than SageMaker ($0.12/hour). It also includes hallucination monitoring tools and a real-time cost dashboard, which could attract mid-sized companies seeking cost control.

What Readers Should Know

• Inkling is not the most powerful model, but it is the most honest: it includes a mechanism to say 'I don't know'.
• Its token efficiency (1/3 the cost of Nvidia Nemotron 3 Ultra on code) makes it attractive for companies with large inference volumes.
• It is open-weight, allowing download and modification, but the business is in the Tinker platform.
• Mira Murati, former CTO of OpenAI, leads this move prioritizing transparency over raw performance.
• Training used 45 trillion multimodal tokens, giving it native capabilities in image, audio, and video.
• The model has a 128k token context window, surpassing Llama 4 (100k) and GPT-5 (96k).
• Early third-party tests show Inkling reduces hallucinations by 40% compared to comparable models, according to an independent analysis from Stanford University.
• Training infrastructure includes 10,000 GB300 NVL72 GPUs, with an estimated cost of $300 million.

'Inkling is not the most powerful model available today, closed or open' — statement from Thinking Machines Lab that breaks with the industry's tradition of hype.

In summary, Inkling represents a paradigm shift: from the race for absolute performance to the pursuit of efficiency, transparency, and control. If companies adopt this approach, it could accelerate the democratization of AI and force giants to rethink their strategies. However, time will tell whether Murati's bet on honesty and openness translates into commercial success or remains an idealistic experiment.