Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, released the full weights for its first foundation model, Inkling, on July 15 under an Apache 2.0 license. By publishing the weights before offering a closed API—and acknowledging that Inkling is not the strongest available model—the 17-month-old laboratory has placed model ownership and deployment control at the center of its first major release.

Inkling is a Mixture-of-Experts model with 975 billion total parameters and 41 billion active parameters. Launch materials summarized by Thorsten Meyer AI say it has a one-million-token context window and was pretrained on 45 trillion tokens spanning text, images, audio and video. It accepts text, image and audio inputs and produces text.

The laboratory published BF16 and NVFP4 checkpoints on Hugging Face with day-one support for transformers, vLLM, SGLang and llama.cpp. The Apache 2.0 license generally permits downloading, modifying, self-hosting and commercial use, although organizations must review all accompanying terms before deployment.

Thinking Machines said Inkling does not lead every open or closed model. Its published results include 97.1% on AIME 2026 and 87.2% on GPQA Diamond, while the model trailed cited competitors on several software-engineering and agent benchmarks. These are vendor-published historical results, some involving a prerelease checkpoint, and await independent replication.

Open Weights Shift Model Control

Publishing weights first changes what customers receive. A closed API provides access to a provider’s service; downloadable weights allow self-hosting, private fine-tuning and deployment without depending on continued API access. That distinction matters to governments, researchers and companies concerned about data control, service availability and vendor dependence.

The weights do not reveal a readable account of how Inkling reaches each answer. They contain learned numerical parameters, not a transparent map of concepts or a complete reasoning record. Because the training dataset and pipeline were not published, the release offers operational control without full reproducibility.

Thinking Machines also introduced an adjustable reasoning-effort setting from 0.2 to 0.99. The company says this lets operators trade reasoning tokens against cost and latency. That claim could make deployment efficiency more relevant than a single peak benchmark score, but real-world cost comparisons have yet to be independently verified.

Murati’s Lab Chooses Openness

Murati founded Thinking Machines after serving as OpenAI’s chief technology officer, and the laboratory employs researchers who worked on ChatGPT. Its first release arrives as developers weigh proprietary hosted systems against open-weight models from American and Chinese laboratories.

Thorsten Meyer AI described Inkling as a Western open-weight alternative, while reporting that GLM-5.2 remains ahead on some reasoning and agent tasks and Kimi K2.6 is competitive in multimodal work. The report also says Inkling’s post-training used synthetic data from Kimi K2.5, showing that model development crosses national and corporate boundaries.

“Inkling is not the strongest model available today, closed or open.”

— Thinking Machines Lab, in its Inkling announcement

Testing and Usage Terms Unsettled

It is not yet clear how Inkling will perform across independent evaluations or production workloads. The reported benchmark figures come from the developer or cited evaluation services, and some results used a prerelease checkpoint. Claims about reasoning efficiency and spontaneous compression of chain-of-thought output during reinforcement learning also need outside verification.

A separate Model Acceptable Use Policy was reported to cover the original parameters and modified versions, including restrictions involving surveillance, deception and automated decisions affecting rights. Thorsten Meyer AI said it had not verified that policy. Until the governing documents are checked directly, the full legal scope of commercial use remains uncertain.

Independent Tests and Smaller Weights

Researchers and prospective users will next test Inkling against competing models on their own data, hardware and latency targets. Running the flagship is expensive: the source estimates that BF16 deployment requires at least two terabytes of aggregate VRAM, while NVFP4 still needs roughly 600 gigabytes.

Attention will also turn to Inkling-Small, a preview model with 276 billion total parameters and 12 billion active parameters. Thinking Machines has said its full weights will follow after testing. That release may determine whether the laboratory’s open-first strategy reaches smaller operators rather than mainly well-funded institutions.

Key Questions

What exactly did Thinking Machines release?

The laboratory released Inkling’s full BF16 and NVFP4 model weights on Hugging Face, together with support for several common inference frameworks. It did not publish the training dataset or complete training pipeline.

Does access to the weights reveal how Inkling thinks?

No. The weights expose the model’s learned numerical parameters, which researchers can inspect and modify, but they do not provide a plain-language record of reasoning or explain how individual facts and behaviors were learned.

Can Inkling be used commercially?

The stated Apache 2.0 license permits commercial use, subject to its conditions. A separate acceptable-use policy has also been reported but was not verified in the supplied source, so prospective users should examine the current model documentation before deployment. This is not legal advice.

Can Inkling run on a personal workstation?

Generally not in its standard released forms. Reported requirements are at least two terabytes of VRAM for BF16 or about 600 gigabytes for NVFP4. Heavily compressed versions may reduce that requirement, while the planned Inkling-Small release could be more accessible.

Is Inkling the best-performing open model?

Thinking Machines says it is not the strongest model across the board. Inkling posts high developer-reported results on several reasoning and audio tests but trails cited competitors on some coding, agent and multimodal evaluations. Independent testing is still pending, and past benchmark results do not guarantee future performance.

Source: Thorsten Meyer AI