Thinking Machines Ships Inkling, Its First Open-Weight Model

The interesting thing about the first model out of Mira Murati's Thinking Machines Lab, a mixture-of-experts system called Inkling, is what the company chose not to claim. In its own release the startup states that Inkling is "not the strongest overall model available today, open or closed." That is an unusual way to launch a first model from a lab that closed what has been reported as one of the largest seed rounds in venture history, $2 billion at a $12 billion valuation, on the promise of building serious AI.

The specs are still substantial. Inkling has 975 billion total parameters with about 41 billion active for any given task, and it was trained on 45 trillion tokens of text, image, audio, and video, reasoning natively across all four. On one narrow measure, the company says Inkling reaches the same coding score as Nvidia's latest open-weight Nemotron 3 Ultra using roughly a third as many tokens. Reported benchmarks include 77.6% on SWE-bench Verified and 97.1% on AIME 2026, according to TechCrunch's write-up of the release.

The strategic frame matters more than the numbers. Inkling is open-weight, so developers can download and customize it without seeing the training data or source code, and Thinking Machines says it is not aiming to monetize the model itself. Revenue is coming from Tinker, a developer tool for fine-tuning that already counts hedge fund Bridgewater Associates among its customers. In effect, Murati's team is giving away the base and selling the picks and shovels for adapting it, a posture backed by Andreessen Horowitz, Nvidia, AMD, Cisco, and Jane Street.

The honest caveat is that a hedged launch does not tell you how the model will hold up in production against closed frontier systems from OpenAI, Anthropic, or Google. What the reporting does not give you is any detail on license terms, per-token inference cost, or safety evaluations, and there is no roadmap beyond Inkling and a smaller preview called Inkling-Small at 276 billion parameters.

If the bet pays, the beneficiaries are the enterprises and US developers who wanted a large, multimodal open alternative to the strongest Chinese open models without waiting for the next flagship drop from Meta.