Thinking Machines ships Inkling, a giant open-weight model

TL;DR:

• Thinking Machines, founded by former OpenAI chief technology officer Mira Murati, has released Inkling, its first general-purpose model.
• Inkling is open-weight — users can download, run and customise it — and at 975 billion parameters is among the largest models of its kind.
• Benchmarks show closed models from Anthropic, Google and OpenAI keep the overall edge, but Inkling is competitive on agent-related tasks.

Thinking Machines has released its first general-purpose model, and it has taken a side in one of the industry’s defining splits. Inkling is open-weight: anyone can download, run and customise it, in contrast to the closed systems sold by the major US labs.

That choice matters because the Western open-model ecosystem has thinned out. Meta retreated to a proprietary approach after the disappointing open release of Llama 4 last year, leaving Chinese labs — Alibaba’s Qwen among them — as the default open alternative for businesses unwilling to pay closed-model prices. Hedge fund Bridgewater Associates used Thinking Machines’ Tinker customisation product to build its own version of Qwen, which it said beat top proprietary models at lower cost. Inkling gives firms like that a Western-origin base model option for the first time in a while.

At 975 billion parameters, Inkling is among the largest open-weight models released. The company’s published benchmarks are candid: closed models from Anthropic, Google and OpenAI keep the overall performance edge, but Inkling puts in a competitive showing on agent-related tasks — the segment where enterprise demand is growing fastest.

Why UK firms should care

For UK organisations, open-weight models carry a specific appeal: they can be run on-premises or in UK data centres, keeping sensitive data out of third-party APIs — a live concern for regulated sectors and the public bodies now repatriating workloads. Provenance also matters to some buyers; procurement teams that hesitate over Chinese-origin models now have a serious US-built alternative at the frontier scale.

Looking forward

Inkling is available through Tinker and other developer platforms. The test is adoption: whether enterprises that standardised on Qwen and its peers will switch for provenance and agent performance, and whether Thinking Machines — valued at around $50 billion — can sustain open releases where Meta could not.