Mira Murati’s Thinking Machines releases 975B-parameter open model

Thinking Machines Lab, the AI company founded in early 2025 by former OpenAI CTO Mira Murati, released its first model on Wednesday: a 975 billion-parameter open-weights system codenamed Inkling. The release matters because there are still few U.S.-built open-weights models at this scale, with much of the recent activity in frontier open models coming from Chinese labs.

Thinking Machines says Inkling is the largest American open-weights model released to date. The company is positioning it against large Chinese systems including DeepSeek V4, GLM 5.2 and Kimi K2.6, while its own benchmark charts show it behind proprietary models from OpenAI and Anthropic. As usual with model benchmarks, the claims should be treated as directional rather than definitive.

The model is being distributed under the Apache 2.0 license, allowing users to modify and fine-tune it for commercial or internal use. Thinking Machines is also offering access through Tinker, its platform for model customization and fine-tuning. The company did not disclose pricing for API access.

Large model, large hardware bill

Inkling requires more than 2 terabytes of GPU memory to run at native 16-bit precision, according to Thinking Machines. That puts the hardware requirement at roughly eight Nvidia B300 accelerators or 16 H200s. For customers without that footprint, the company also released an NVFP4 quantized version that it says can run on about half as many GPUs.

Thinking Machines describes Inkling as a model for developers building AI applications, while also saying it can handle general uses such as chatbots. The company says the model can generate fine-tuning scripts to adjust its behavior, learn new capabilities and test its own performance. Those are company claims, and the release does not provide independent validation of how well those workflows perform in production.

The model supports a context window of 1 million tokens, which could make it useful for large-codebase work and long-document retrieval tasks. Long context windows have become table stakes for vendors selling into software and enterprise search use cases, but real-world usefulness depends heavily on retrieval quality, inference cost and latency.

DeepSeek influence, Nvidia training stack

Thinking Machines says Inkling uses a mixture-of-experts architecture inspired by DeepSeek-V3, but that it trained the model from scratch on Nvidia GB300 NVL72 systems. The training data mix, according to the company, covered 45 trillion tokens across text, images, audio and video.

The model includes 256 routed experts and two shared experts. For each generated token, six experts are active, representing about 41 billion parameters. Thinking Machines says that design should allow Inkling to produce tokens at a rate similar to DeepSeek V4 when both are run on comparable hardware.

Inkling is also a reasoning model, meaning it was trained with reinforcement learning to use chain-of-thought-style intermediate tokens before producing answers. Thinking Machines says it tuned the model to use fewer of those reasoning tokens, claiming that Inkling matches Nvidia’s Nemotron 3 Ultra on Terminal Bench 2.1 while using about one-third as many tokens. That claim matters for cost, since reasoning tokens are still billable tokens in API settings.

Inkling is available through Tinker and for download through model repositories including Hugging Face. Thinking Machines says it is working to make the model available through third-party services including TogetherAI, Fireworks, Modal, Databricks and Baseten. At launch, the company claims support for inference engines including vLLM, SGLang, Miles, TokenSpeed and Llama.cpp.

The company is also previewing Inkling-Small, a 276 billion-parameter mixture-of-experts model with 12 billion active parameters. Thinking Machines says the smaller model is aimed at users prioritizing latency over throughput and quality, and that it plans to release the weights after testing is complete.

This story draws on original reporting from The Register.