Inkling AI Model: Thinking Machines Lab Launches 975B Open-Source Multimodal

What happened

Thinking Machines Lab has released Inkling, its debut AI model, marking the Philippine-founded startup's first public entry into the competitive large-model landscape. Inkling is a 975-billion-parameter open-source model built with multimodal capabilities, trained to process and reason across video and audio in addition to text — a design choice that distinguishes it from many text-first rivals.

The release positions Thinking Machines Lab as a challenger to established players such as Anthropic and OpenAI, with the open-source licensing intended to accelerate adoption and community development. By making Inkling openly available, the company is betting that transparency and accessibility will help it carve out credibility in a field dominated by well-capitalised incumbents.

Why it matters

For customer experience and service design practitioners, the arrival of a capable open-source multimodal model is significant. The ability to process video and audio natively — rather than relying on text transcriptions or image stills — opens more faithful channels for analysing real customer interactions: call recordings, video support sessions, in-store footage and voice-of-customer data captured in its original, uncompressed form. That fidelity matters enormously when the goal is to understand not just what customers say, but how they say it.

From a behavioural economics standpoint, tone, hesitation, facial expression and ambient context are rich signals of emotional state and decision friction. Models trained on video and audio can, in principle, surface those signals at scale — enabling service teams to identify moments of confusion, frustration or delight that text-based analytics routinely miss. Open-source access lowers the barrier for CX teams and service-design researchers to experiment with these capabilities without enterprise licensing costs.

By the numbers

• 975 billion parameters — the scale of Inkling, Thinking Machines Lab's debut model, placing it among the largest publicly released open-source models.
• 3 modalities — text, video and audio are all supported natively within a single model architecture.

The Renascence take

Most commentary on Inkling will focus on the parameter count and the competitive optics against OpenAI and Anthropic. That framing misses the more consequential detail for anyone running customer operations: this is one of the first openly available models trained to understand the texture of human communication, not merely its transcript.

The CX industry has spent a decade mining text — surveys, chat logs, reviews — while the richest behavioural signals have sat locked inside audio and video that was too expensive or too complex to analyse at scale. Inkling's open-source multimodal architecture changes that calculus. The organisations that will benefit most are not those racing to deploy it fastest, but those disciplined enough to define the specific emotional and behavioural moments they want to detect before they touch the tooling. Model capability without a service-design hypothesis is just noise at 975 billion parameters.