Inkling AI Model: Mira Murati's US Open-Weight Challenger for Enterprises

What happened

Thinking Machines Lab, the San Francisco-based startup founded by former OpenAI chief technology officer Mira Murati, has released Inkling — its first general-purpose AI model. The launch positions the company as a US-headquartered contender in the open-weight AI space, a segment where Chinese developers have held several of the leading positions in coding and reasoning benchmarks.

Inkling is built on a mixture-of-experts architecture, carrying 975 billion total parameters while activating only 41 billion during any given inference pass — a design choice that balances raw capability with computational efficiency. The model supports a context window of up to one million tokens and was pretrained on 45 trillion tokens drawn from text, images, audio, and video. Thinking Machines also trained Inkling specifically for coding, tool use, and multimodal tasks.

The release builds on the company's October 2025 launch of Tinker, an API-based platform that allows developers and enterprises to customise AI models. Inkling can be fine-tuned directly through Tinker, giving organisations a route to adapt the model for domain-specific applications. The move comes as AI model routing platform OpenRouter, in a June 2026 assessment, spotlighted a cluster of Chinese-developed models — including DeepSeek V4 Flash, GLM 5.2, MiniMax M3, and Nvidia Nemotron 3 Ultra — as standout performers, underscoring the competitive pressure that Inkling is designed to address.

Why it matters

For organisations designing and operating customer-facing services, the provenance of an AI model is no longer a purely technical concern — it is a procurement, compliance, and trust question. Enterprises in regulated industries and government-adjacent sectors face growing scrutiny over where their AI infrastructure originates, who controls the weights, and what data sovereignty obligations apply. A capable, openly available US-developed model with a clear customisation pathway reduces one layer of that friction.

From a service-design perspective, Inkling's multimodal pretraining and million-token context window are the more immediately consequential features. Long-context models can hold an entire customer history, a complex policy document, or a multi-session conversation in a single inference pass — reducing the fragmentation that degrades service quality when agents or automated systems lose thread across interactions. The ability to fine-tune through Tinker means organisations can shape the model's behaviour around their specific service protocols rather than working around a fixed, general-purpose personality.

By the numbers

• 975 billion total parameters in the Inkling mixture-of-experts model
• 41 billion parameters active during each inference pass
• 1 million tokens maximum supported context window
• 45 trillion tokens of multimodal data used in pretraining, spanning text, images, audio, and video
• October 2025 — launch date of Tinker, Thinking Machines' earlier API customisation platform

The Renascence take

Most coverage of Inkling will focus on the geopolitical framing — US versus Chinese open-weight models — and miss the more operationally significant detail: the combination of a very long context window with a fine-tuning platform is what actually changes the calculus for service operators, not the parameter count or the founder's biography.

The real CX opportunity here is not "which country made the model" but whether organisations will use the fine-tuning capability to encode genuine service standards — empathy heuristics, escalation logic, brand tone — rather than simply deploying a general model with a thin system prompt. Behavioural economics tells us that consistency and predictability are foundational to customer trust; a fine-tuned model that reliably behaves within defined service boundaries will outperform a more powerful but unpredictable one every time. Customer-obsessed operators should treat Tinker not as a developer tool but as a service-design instrument — and invest the same rigour in defining the model's behaviour as they would in training a human frontline team.