Thinking Machines launches Inkling, a 975B open-weight model built for customization

Thinking Machines Lab released Inkling on July 15 as its first in-house model and its first open-weights release. The model is a sparse Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active for each token, and a context window of up to 1 million tokens. It was pretrained on 45 trillion tokens spanning text, images, audio, and video. Inkling accepts text, image, and audio inputs, but its output is text, including code and structured artifacts. 1 2

The weights are released under Apache 2.0 and are downloadable from Hugging Face. Developers can also access Inkling through Tinker for fine-tuning or through third-party inference providers. That makes the launch more than an API announcement: the intended workflow is to start with a broad base model, then adapt it to a company or product's own data and behavior. 2

Thinking Machines reports Inkling at a high thinking-effort setting with 77.6% on SWE-bench Verified, 63.8% on Terminal Bench 2.1, 73.5% on MMMU Pro, and 91.4% on VoiceBench. The company also says Inkling can reach the same Terminal Bench performance as Nvidia's Nemotron 3 Ultra with roughly one-third as many generated tokens, a claim that points to the model's controllable effort setting as much as to its raw capability. 2

Those comparisons are vendor-reported, not an independent evaluation. The model card notes that some Inkling results use internal harnesses, while external models may use self-reported numbers. Thinking Machines also says plainly that Inkling is not the strongest overall model available today. Its pitch is breadth, multimodal input, and efficiency for customization rather than a clean leaderboard win. 1

Inkling is a concrete test of Mira Murati's startup thesis that organizations should be able to shape their own models instead of renting a fixed general-purpose assistant. TechCrunch describes the release as a bet against one-size-fits-all AI, with Tinker as the service layer around the downloadable model. That creates a different competitive target from ChatGPT, Claude, or Gemini: the product is the adaptation loop, not only the base model. 3

The trade-off is substantial. The official card lists at least 2 TB of aggregate VRAM for the BF16 checkpoint, or about 600 GB for the NVFP4 option. It also warns about hallucinations, instruction failures, weaker performance in long conversations, training-cutoff limits, and occasional compliance with indirectly framed harmful prompts. Inkling is permissively licensed and adaptable, but it is not a lightweight local model or a drop-in production assistant. 2