Thinking Machines Lab Launches Inkling to Challenge the Industry Focus on Monolithic AI Scale

The Historic Rise and Customization Engine

Thinking Machines Lab, co-founded by former OpenAI chief technology officer Mira Murati and researcher Lilian Weng, entered the industry with unprecedented financial backing. The startup secured a historic 2 billion dollar seed round at a 12 billion dollar valuation, marking the largest seed funding in venture capital history.

This massive capital injection, led by Andreessen Horowitz, occurred when the company was only six months old and had yet to ship a product. Shortly after, in October 2025, the company launched its first developer-facing infrastructure product called Tinker.

Tinker allows developers to write Python code locally and run it on distributed GPU clusters via an API. Former OpenAI researcher Andrej Karpathy described Tinker as a clever way to slice post-training complexity, removing 90 percent of infrastructure headaches while keeping 90 percent of algorithmic control.

Tyler Griggs from the SkyRL group at Berkeley noted that Tinker allows researchers to focus on actual research rather than engineering overhead. Amidst this rapid growth, reports surfaced in October 2025 that a co-founder had left the startup to join Meta in a deal valued at 3.5 billion dollars in liquid stock.

Challenging the Scaling Orthodoxy with System 2 AI

Rather than simply chasing larger models, Thinking Machines Lab is betting heavily on System 2 AI, which prioritizes real-world reasoning, self-correction, and chain-of-thought prompting. This approach stands in contrast to the speed and pattern recognition of System 1 AI, which dominates current commercial models.

At the TED AI conference in San Francisco, reinforcement learning researcher Rafael Rafailov argued that the first true superintelligence will be a superhuman learner rather than just a scaled-up system. Rafailov stated that learning is something an intelligent being does, whereas training is merely something being done to it.

The company believes that current AI systems fail because they cannot learn dynamically from their ongoing experiences. This philosophy underpins their focus on continuous human-AI collaboration, rejecting the idea that humans should simply hand off tasks to autonomous models and walk away.

Inkling and the Shift to Real-Time Interaction

On July 15, 2026, Thinking Machines Lab launched its first open-weight proprietary model, named Inkling. Inkling is a mixture-of-experts model containing 975 billion total parameters, though it only utilizes roughly 41 billion active parameters for any single task to maintain speed and cost efficiency.

Trained on 45 trillion tokens of text, images, audio, and video, Inkling is capable of reasoning natively across all of these modalities simultaneously. The model is designed to support the startup's bet on highly customizable, open AI over one-size-fits-all proprietary systems.

Inkling's release follows a May research preview of the company's Interaction Models, which aim to reform how voice AI operates. Traditional voice systems like OpenAI's GPT-Realtime-2 and Google's Gemini Live rely on external helper systems, such as voice activity detectors, which freeze the model's perception while it speaks.

Thinking Machines Lab addresses this bottleneck by processing audio, video, and text streams in parallel 200-millisecond chunks, known as time-aligned micro-turns. This allows the system to handle interruptions, react to visual cues, and speak simultaneously, outperforming traditional turn-based setups on interaction quality.

Whether Thinking Machines Lab can successfully leverage its massive treasury to prove that open-weight, highly customizable reasoning engines can outpace the sheer scale of centralized tech giants remains the industry's most expensive question.