Thinking Machines Releases Inkling - a 975B Open-Weight Model That Admits It's 'Not the Strongest'

Thinking Machines Lab, the AI startup led by former OpenAI CTO Mira Murati, released its first in-house model “Inkling” on July 15, 2026¹. It is a Mixture-of-Experts model with 975 billion total parameters, distributed as open weights that anyone can download and modify, under the commercially permissive Apache 2.0 license².

What stands out is the company’s own framing: the official blog states plainly that “Inkling is not the strongest overall model available today, open or closed”¹. For a lab that spent roughly a year and a half building AI infrastructure largely out of public view, choosing a debut that explicitly does not chase the performance crown signals a strategy quite different from the major labs.

Inkling’s Specs - 975B Parameters and a 1M-Token Context Window

Inkling is a 66-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) feed-forward backbone². While the total parameter count is 975B, only 41B parameters are active during inference, which keeps the model relatively light to run for its size — the core advantage of MoE architectures¹.

The model was pretrained on 45 trillion tokens of text, images, audio and video¹. It accepts text, image, and audio inputs, and currently outputs text only². The context window extends up to 1 million tokens¹.

Inkling also offers controllable “thinking effort,” letting users trade performance against token consumption per task¹. According to the official blog, it matches competing models on the Terminal Bench 2.1 coding benchmark while using roughly a third of the tokens¹. TechCrunch identifies the comparison target as NVIDIA’s Nemotron 3 Ultra³.

A lighter variant, “Inkling-Small,” with 12B active parameters, is also available in preview¹.

Availability - Weights on Hugging Face, but Serious VRAM Required

Inkling’s weights are published on Hugging Face, and the model is also available through API providers including TogetherAI, Fireworks, Modal, Databricks, and Baseten¹. Running it on your own infrastructure takes 2TB of VRAM for the BF16 checkpoint and 600GB even for the NVFP4 quantized version⁵ — well beyond a personal local LLM setup, positioning it squarely for enterprises and research institutions.

Tinker, the company’s fine-tuning developer tool, supports fine-tuning Inkling from day one, with a 50% discount for a limited time¹⁴.

The Strategy of Not Chasing “Strongest” - a Bet on Customizable AI

TechCrunch sums up Thinking Machines’ central bet: AI that organizations can adapt for themselves will outperform the one-size-fits-all models the biggest labs currently sell³. Flagship models from OpenAI, Anthropic, and Google are powerful, but their weights are closed — users take the behavior they are given. Inkling inverts this: it is designed from the start for organizations to modify with their own data and workflows.

The strategy is inseparable from the business model. The company does not monetize the model directly; it earns revenue through Tinker, its fine-tuning developer tool, with customers including the hedge fund Bridgewater Associates⁴. Inkling is effectively the raw material for customization, and Tinker is the paid machinery for shaping it.

In the open-weight arena, Chinese players such as Moonshot’s Kimi K2 have led the way. Fortune frames Inkling’s open-weight approach as addressing the U.S. market gap versus Chinese competitors⁴.

The Backstory - a $12B-Valuation Lab’s First Public Proof Point

Thinking Machines Lab was founded by Murati, who served as CTO at OpenAI, and employs roughly 200 people³. In 2025 it raised $2 billion at a $12 billion valuation⁴. A larger follow-on round was reported to be in the works, but according to TechCrunch it had stalled by January 2026³.

Since its founding, the company had built its AI infrastructure largely out of public view; TechCrunch calls Inkling the company’s first public proof point after a year and a half³. The official blog says “Inkling is just the start: our first release in a model family we will continue to build on”¹, with models of different sizes planned.

Renting AI vs. Growing Your Own

For enterprise AI adoption, renting a frontier-class model through an API has so far been essentially the only practical option. If customization-first open-weight models like Inkling reach production quality, owning a purpose-built model adapted with your own data becomes a realistic alternative — particularly attractive for regulated industries that cannot send data to external APIs, or businesses whose workflows don’t fit one-size-fits-all models.

The hurdles remain real: infrastructure costs starting at 600GB of VRAM, and the talent needed to design and evaluate fine-tuning. As the company’s own candid positioning suggests, flagship APIs still win on general capability. What Inkling tests is not the summit of performance but a different competitive axis — adaptability. Whether the bet pays off will be measurable by how many enterprise customization case studies emerge through Tinker in the coming months.