Mira Murati Surprises Everyone: Her First AI Model Admits It’s Not the Best
Thinking Machines Lab
Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, has announced Inkling, its first internally developed artificial intelligence model. The strategic choice is clear from the debut: instead of claiming absolute leadership in performance rankings, the company proposes an open-weight model designed to be customized by businesses and adapted to specific application scenarios.
Inkling is a Mixture-of-Experts (MoE) model with a total of 975 billion parameters, of which about 41 billion are activated during each inference. It supports a context window of up to 1 million tokens and has been trained on 45 trillion tokens from text, images, audio, and video. The model is able to natively process multimodal content, although it currently exclusively generates textual output, including code and structured data.
Thinking Machines emphasizes that Inkling "is not the most powerful model available today, neither among open nor proprietary models". According to the company, the value of the project lies instead in the balance between general capabilities, efficiency, and ease of customization through the proprietary Tinker platform, which allows for fine-tuning while maintaining ownership of the resulting model.
The stated goal is to enable organizations to build specialized AI systems using their internal knowledge, rather than relying exclusively on closed generalist models. In this vision, the base model represents merely a starting point, while the true competitive advantage derives from adapting to the user company's data and processes.
From a technical perspective, Inkling introduces a system of "controllable thinking effort", allowing developers to dynamically adjust the model's reasoning level, balancing accuracy, latency, and computational cost. According to benchmarks published by the company, on Terminal Bench 2.1, the model achieves performance comparable to NVIDIA's Nemotron 3 Ultra while using about one-third of the generated tokens.
Thinking Machines describes Inkling as a generalist model capable of operating in numerous areas, from mathematical reasoning to agent-based programming, including visual comprehension, audio processing, factual search, and dialogue. In the benchmarks released by the company, the model shows competitive results against other open-weight solutions, while generally falling behind the leading proprietary models in more complex tests.
Among the highlighted features is also the multimodal architecture developed without dedicated encoders for images and audio. Images are converted into 40x40 pixel patches, while audio is represented through dMel spectrograms; both data types are processed alongside textual tokens within the model.
Inkling also integrates mechanisms designed to enhance the reliability of responses. The lab stated it has worked on confidence calibration, adherence to instructions, and reducing hallucinations through a combination of reinforcement learning, automatic claim verifiers, and external evaluation systems. On the security front, the published benchmarks indicate high performance in FORTRESS and StrongREJECT tests, designed to assess the model's ability to reject harmful requests without compromising legitimate ones.
At the same time, Inkling-Small was also introduced, a more compact version based on 276 billion total parameters and 12 billion active ones. While offering slightly lower performance in some benchmarks, the model aims to reduce costs and latency, making it suitable for workloads such as coding, automated assessment via LLM, and synthetic data generation.
According to Thinking Machines, Inkling was developed in about nine months using NVIDIA GB300 NVL72 systems. For the initial training phase, limited distillation activity was also utilized, leveraging existing open-weight models, including Kimi K2.5, while the lab states that future models will be trained entirely with proprietary pipelines.
The model is already available on the Tinker platform with context windows of 64,000 and 256,000 tokens, as well as through numerous partners in the AI ecosystem, including Hugging Face, Together AI, Databricks, Fireworks, Baseten, and Modal. The complete weights are also distributed in NVFP4 format optimized for NVIDIA Blackwell systems.