After almost a year and a half of operating out of public attention, Thinking Machine Labs, the high-profile AI startup founded by former OpenAI Chief Technology Officer Mira Murati, has officially revealed its first AI model.
The startup has unveiled Inkling, its flagship model that is multimodal and open-weight and is built from scratch. Unlike the frontier models of today, Inkling is the latest addition to the rising group of new AI models designed to challenge Silicon Valley with an offline approach.
Inkling, however, is different from most of the frontier AI models available in the market today. Murati’s model allows enterprises to download and run it locally, unlike the subscription-heavy, cloud-dependent frameworks of rivals like OpenAI’s GPT 5.6 or Anthropic’s Claude Fable.
How is Inkling different from GPT
Inkling boasts a massive 975 billion total parameters, with 41 billion active parameters dynamically deployed for any given task. It was pre-trained from scratch on 45 trillion tokens spanning text, images, audio, and video, thus giving it multimodal reasoning capabilities. While it can accept and understand all four formats, its outputs are currently all textual, including complex coding, styled artefacts, and structured data.
Murati’s venture sent shockwaves through the industry in February 2025 when it secured a historic $12 billion seed funding round – one of the largest in tech history. Since then, the 200-employee firm has operated in virtual secrecy, crafting an AI architecture intended to upend the current “one-size-fits-all” status quo.
Alongside Inkling, the startup has also launched Tinker as a dedicated developer tool designed to let users fine-tune Inkling’s weights for bespoke workflows.
Inkling, however, isn’t like ChatGPT or Claude. The model introduces features aimed solely at enterprise concerns such as calibrated answering. Instead of making blind, hallucinated guesses when it lacks data, the model explicitly signals uncertainty. It also introduces a variable “thinking effort” toggle, giving users the freedom to manually balance processing speed against deep-reasoning performance.
“Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today,” Murati announced in a post on X.
Inkling’s performance and efficiency benchmarks
In benchmarking data shared by the company, Inkling trailed behind top-tier closed frontier models like GPT 5.6, Claude Fable 5, and Moonshot AI’s Kimi K2.6. Murati’s startup, however, isn’t making big claims as far as raw performance is concerned.
Where Inkling excels is raw token efficiency. On the Terminal Bench 2.1 benchmark, the model matched the performance of Nvidia’s Nemotron 3 Ultra while requiring roughly a third of the tokens.
Note that Inkling was trained entirely on Nvidia GB300 NVL72 hardware systems, utilising an infrastructure partnership established in March. The company also confirmed that while the model was pre-trained from scratch, its early post-training data partially leveraged outputs from Moonshot AI’s Kimi K2.5 before transitioning to massive-scale reinforcement learning.
TML’s going for versatility
To cater to varying developer needs, Thinking Machine Labs also introduced a preview of Inkling-Small. Operating on 12 billion active parameters, this lighter-weight version is engineered to deliver lower latency and significantly lower operational costs for smaller enterprise workflows. The company stated it has plans to expand the Inkling family line in the future.