Thinking Machines Lab launched Inkling on Wednesday, an artificial intelligence model with 975 billion parameters. The company, founded in early 2025 by former Open AI CTO Mira Murati provides the model under the Apache 2.0 license, which allows developers to tweak and use it freely.
Inkling is currently the largest open weight model in the United States and is positioned as an alternative to Chinese open weight models such as DeepSeek V4, GLM 5.2, and Kimi K2.6. It is now available on Thinking Machines’ Tinker platform and through Hugging Face for direct download.
Model specifications, hardware requirements and positioning against Chinese open weights
Inkling combines a mix of expert architecture with the following features:
- 975 billion total parameters
- 256 routed experts and two shared experts
- Six experts are activated for each token, resulting in around 41 billion parameters during inference. Supports one million symbolic context window.
- The model was trained with 45 billion tokens spanning text, images, audio and video and was developed using Nvidia GB300 NVL72 systems.
While inspired by DeepSeek-V3, Inkling was trained from scratch by Thinking Machines, rather than fine-tuned from existing weights.
Running Inkling at native 16-bit precision requires more than two terabytes of GPU memory. Practical hardware configurations include around eight Nvidia B300 accelerators or sixteen Nvidia H200 accelerators.
For users with less GPU power, Thinking Machines offers an NVFP4 quantized version of the model that requires about half the GPU. This version trades some accuracy for reduced memory usage while maintaining most of the model’s capabilities.
Reasoning capabilities, Apache 2.0 licenses, and developer access
Like most modern frontier models, Inkling is a reasoning model trained with reinforcement learning to use chain-of-thought thinking before responding.
Thinking Machines claims that Inkling matches Nvidia’s Nemotron 3 Ultra, which was previously the largest open-weight US model with 550 billion parameters.
It reportedly achieves comparable results in Terminal Bench 2.1 using about a third of the thought tokens. This efficiency means that Inkling could offer similar performance at a lower cost when the price is token-based.
Users should note that thought tokens are billed like any other token, so longer thought chains may increase the cost of each response. Inkling enters a market where Chinese labs have dominated open-weight AI development. Competitors include:
- DeepSeek V4 and V4-Pro from Huawei Ascend,
- GLM 5.2, and
- Like K2.6.
Thinking Machines claims that Inkling is competitive with these Chinese models in various workloads. However, users are advised to independently verify such claims through their own testing, as gaming industry benchmarks for AI are often unreliable.
Inkling’s benchmark charts also show it behind proprietary models like Anthropic’s Claude and OpenAI’s GPT. The key trade-off for users is getting access to weights and full customization rather than relying on closed models with higher raw performance.
Tuning, automodification, and hotspots for Inkling are also part of a larger discussion.
Thinking Machines describes Inkling as highly adaptable for developers creating AI applications and general-purpose uses like chatbots. The Apache 2.0 License permits commercial use, modification, and redistribution.
The company also claims that Inkling can write its own tuning scripts to refine its behavior, learn new skills, and evaluate its capabilities. The self-modification capability aims to make customization more accessible to developers without deep machine learning experience.
The company’s Tinker platform provides customization and tuning tools. For developers who want to use Inkling:
- Tinker Platform – Available now for API access and tuning tools
- Hugging Face: Model weights available for direct download
- Third-party API providers: Thinking Machines is working with TogetherAI, Fireworks, Modal, Databricks, and Baseten to add the model to their services.
Inkling supports a wide range of inference engines at launch, including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp.
What users should do, Inkling-Small preview and availability
For developers interested in Inkling for their projects: Review the comparative performance of the model relative to your specific use case rather than relying solely on general statements. Consider whether the 1 million token context window justifies the hardware requirements for your workload.
If full precision hardware is not possible, consider the quantized NVFP4 version. Test the model through the Tinker platform before committing to a self-hosted deployment. It is also useful to compare token efficiency claims with real-world workloads in your domain.
For users exploring Inkling for hobby or evaluation purposes: download the model via Hugging Face. Configure an inference engine such as Llama.cpp for local testing. Be prepared for significant hardware requirements, even with quantization.
In addition to Inkling, Thinking Machines is previewing Inkling-Small, a 276 billion parameter combination of experts (MoE) model with 12 billion active parameters.
This smaller model is aimed at users who prioritize low latency over performance and quality. It hasn’t been released yet, but the company plans to release the weights once testing is complete. No specific timeline for the release has been provided.
The launch of Inkling comes at a time when the frontier AI market is increasingly dominated by proprietary models with limited access. OpenAI recently postponed the public release of GPT-5.6 at the request of the US government.
anthropic Fable 5 and Mythos 5 were temporarily suspended before being restored. Alibaba also restricted access to Claude Code for employees for tracking purposes.
In this context, Thinking Machines’ decision to release Inkling under the Apache 2.0 license offers an alternative for developers who want to access frontier-scale capabilities without relying on a single vendor. The name refers to the fictional supercomputer manufacturer from Jurassic Park.
Inkling is now available on the Tinker and Hugging Face platform. Access to third-party APIs through platforms such as TogetherAI, Fireworks, Modal, Databricks and Baseten is expected to follow.
The company has not yet announced pricing for the Tinker platform. Interested users can follow the company’s official channels for updates on Inkling-Small and future releases.
