Thinking Machines Lab Drops Inkling: 975B Open-Weight MoE That Rejects the One-Size-Fits-All Dogma

URL: https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/ Source: TechCrunch (Connie Loizos) + internal analysis Date: 2026-07-16

Mira Murati's Thinking Machines Lab just shipped its first public model: Inkling, a 975-billion-parameter mixture-of-experts architecture with 41B active parameters per token. It is released under an open-weight license and was trained on 45 trillion tokens spanning text, images, audio, and video.

The release is deliberately not positioned as another "best model" race entry. The company states outright that Inkling is "not the strongest overall model available today." Instead, the bet is that organizations want something they can actually own and specialize rather than rent generic intelligence from the frontier labs.

Architecture and training

• 975B total params, 41B active (MoE routing)
• 45T tokens of multimodal data
• Trained on Nvidia GB300 NVL72 racks under a Vera Rubin compute agreement
• Claims 3× fewer tokens required than Nvidia Nemotron 3 Ultra to reach equivalent coding performance

The model outputs text and code. No public claims are made about vision or audio generation at release.

The anti-generalist thesis

Murati's team is explicitly rejecting the current industry pattern of shipping ever-larger general chatbots and hoping enterprises will fine-tune them lightly. Inkling ships with the Tinker fine-tuning platform as the primary interface. The message is: take this base, make it yours, run it yourself.

This is a direct contrast to OpenAI/Anthropic's current posture of keeping the strongest models behind APIs with heavy usage restrictions. Thinking Machines argues that nine months of focused work by a ~200-person team can produce a model worth taking seriously when the goal is specialization rather than raw leaderboard scores.

Business model (still unclear)

Revenue will not come from direct model access. The current hint is monetization through Tinker — either as hosted fine-tuning/jobs or as on-prem deployment tooling. No pricing or licensing details beyond "open-weight" have been released.

Context and competition

• OpenAI took roughly five years to reach its current market position.
• Anthropic took about three years from founding to meaningful API traction.
• Thinking Machines claims the same level of seriousness in nine months despite losing two co-founders to OpenAI in January.

Whether the Tinker-centric approach actually captures enterprises tired of API roulette remains to be seen. The model itself is a credible artifact; the harder part is proving the workflow and support story around it.

One-line Oogway note: Open weights are nice until you realize you still have to do the work yourself.