Thinking Machines Debuts Inkling as Open-Weight Mixture-of-Experts Model

Thinking Machines Lab, the artificial intelligence startup established by former OpenAI Chief Technology Officer Mira Murati, released its inaugural open-weight foundation model on July 16, 2026. Known as Inkling, the architecture utilizes a mixture-of-experts transformer design featuring 975 billion total parameters and 41 billion active parameters.

The model supports a context window reaching one million tokens and underwent pre-training on a dataset of 45 trillion tokens encompassing text, images, audio, and video. According to the company, this massive training scale was intended to provide a versatile base for various multimodal applications across diverse enterprise environments.

Alongside the primary release, the company introduced Inkling-Small, a scaled-down version containing 276 billion total parameters and 12 billion active parameters. This smaller variant is designed to offer a more accessible entry point for developers with limited compute resources who still require high-quality reasoning capabilities.

Development teams can access the model via the proprietary Tinker platform, which provides specific context options for fine-tuning at 64,000 and 256,000 tokens. Thinking Machines has also published full model weights on Hugging Face, including an NVFP4 checkpoint specifically engineered for NVIDIA Blackwell hardware systems to optimize throughput and memory efficiency.

Inference support extends across multiple frameworks, including SGLang, vLLM, TokenSpeed, llama.cpp, and Hugging Face Transformers. The company has integrated the model into third-party hosting environments such as Together AI, Fireworks, Modal, Databricks, and Baseten to facilitate broader deployment for researchers and engineers.

Technical specifications indicate that Inkling utilizes a reinforcement learning process involving more than 30 million rollouts during the post-training phase. This training regimen was conducted on NVIDIA GB300 NVL72 systems, resulting in a model that produces more concise reasoning traces without compromising final output accuracy or factual consistency.

The architecture handles multimodal inputs natively without requiring dedicated encoders for image or audio processing. Capabilities include speech transcription, chart analysis, and the execution of Python tools for image manipulation tasks such as zooming and cropping, which enhances the agentic utility of the model.

Thinking Machines explicitly positions Inkling as a base for customization rather than a pursuit of top-tier benchmark dominance. As stated in a company blog post, the model is intended to serve as a foundation for developers to make it their own.

Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customisation: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.

Performance metrics on Terminal Bench 2.1 show Inkling achieving results comparable to NVIDIA’s Nemotron 3 Ultra while generating approximately one-third of the total token volume. While it trails models like GLM 5.2 and Kimi K2.6, it demonstrates competitive standing on agentic coding benchmarks and the Design Arena Agentic Web Development leaderboard.

The focus on factual calibration and instruction following suggests a strategic effort to balance utility with safety. External testing indicates the model maintains a high threshold for refusing harmful requests while minimizing the occurrence of unnecessary refusals for benign user prompts, which is essential for production-grade applications.

Future adoption will likely depend on how effectively the Tinker platform lowers the barrier to entry for specialized domain training. As the open-weights ecosystem continues to expand, the ability to adjust reasoning effort to manage the trade-off between inference cost and performance will remain a critical watchpoint for data scientists and MLOps engineers.

The company has also launched an Inkling Playground for developers to test these reasoning capabilities in real-time. This interface allows users to observe how the model handles complex, multi-step queries across different modalities in a controlled environment.

Looking ahead, the industry will monitor how the 50% introductory discount on the Tinker platform influences adoption rates among enterprise users. The long-term viability of the model will be tested by its ability to maintain performance consistency across diverse, real-world deployment scenarios and evolving safety requirements.