Thinking Machines Lab launches Inkling 975B parametric model inspired by DeepSeek

Machine Learning


TL;DR

  • License and access: Although Apache 2.0 terminology has been applied to Inkling, Thinking Machines Lab plans to publish Inkling-Small weights after testing.
  • Sparse design: The 975 billion parameter model enables 41 billion parameters per token through an expert mix architecture inspired by DeepSeek-V3.
  • Hardware floor: Inkling’s 16-bit weight requires at least 2 TB of GPU memory, while the lower precision version requires 600 GB.
  • Recruitment test: Various benchmark results, paid Tinker tweaks, and managed hosting will shape Inkling’s real-world adoption.

Thinking Machines Lab released Inkling under the Apache 2.0 license on July 15th after using Kim K2.5 for initial post-training data. Openweight allows developers to download and adapt trained parameters, but the training data and full development process remain private.

Inkling has 41 billion active parameters, totaling 975 billion per token. The design clearly follows DeepSeek-V3, although its scale provides US developers with another large-scale model that can be customized.

Inkling’s mixed expert architecture routes each token through 6 of the 256 experts and 2 shared experts, rather than activating the entire network. Selective routing reduces the computation used per token and explains how a model with 975 billion total parameters uses a much smaller active set. DeepSeek introduced its predecessor architecture in late 2024.

What you need to run and customize Inkling

Sparse routing does not remove hardware barriers. Inkling’s full 16-bit checkpoint requires a GPU cluster with a minimum of 2 TB of total GPU memory. This equates to 8 Nvidia B300 graphics processing units or 16 H200 units. The lower precision version reduces the minimum capacity to 600 GB and can run on 4 B300 or 8 H200 units, reducing memory demands without establishing the same precision.

Local operations still require an inference deployment framework such as SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face, which adds software integration effort to the hardware bill. Their small footprint exceeds that of a typical workstation, leaving self-hosting to organizations with expensive multi-accelerator systems.

Thinking Machines Lab's Inkling Benchmark

Inkling accepts text, images, and audio as input, but returns only text for coding, tool usage, and search extension generation systems. Thinking Machines Lab trained Inkling from scratch on an Nvidia GB300 NVL72 system using 45 trillion training tokens across text, images, audio, and video, but the scale of the training does not guarantee reliable answers.

Developers can download weights from Hugging Face, fine-tune Inkling through Thinking Machines Lab’s paid Tinker platform, or use Inkling Playground. While Databricks made its models available through the Unity AI Gateway for enterprise data, agents, and coding workflows, third-party inference providers provided another path for teams without local hardware. Downloading weights gives adopters more control, but managed access changes who operates the accelerator without reducing Inkling’s underlying memory demands.