AI Startup Founded by Mira Murati Releases First Open Model 'Inkling'

Thinking Machines, an AI startup founded by former OpenAI CTO Mira Murati, has released 'Inkling,' its first open-source language model. The model is offered under an Apache 2.0 license suitable for commercial use, and model weights are already available from Hugging Face and the company's proprietary model training API 'Tinker.'

Inkling is a multimodal model capable of processing text, images, and audio, with a total of 975 billion parameters. The architecture adopts a design called 'Mixture-of-Experts (MoE),' which dynamically utilizes only the necessary portions of all parameters, with active parameters reaching 41 billion. A lighter variant, 'Inkling-Small' (with 276 billion total parameters), was also previewed. This is optimized for workloads prioritizing low latency and low cost.

Examining benchmark performance, in the software engineering domain, Inkling achieved 77.6% on 'SWE-bench Verified,' surpassing the 71.9% of Nvidia Nemotron 3, another open-weight model. On 'VoiceBench' for voice understanding, it achieved 91.4%. However, on the frontend web design metric 'Design Arena Agentic Web Dev,' it scored 1,257 points, placing it in the mid-to-upper range of high performance, while falling short of the coding and reasoning-specialized performance demonstrated by Chinese-origin models like GLM 5.2 (62.1% vs. Inkling's 54.3% on SWEBench Pro, 82.7 vs. 63.8 on Terminal Bench 2.1, among others). Thinking Machines positions its model not as 'specialized in specific domains' but as a 'generalist responding to a wide range of applications.'

One of Inkling's distinctive features is the mechanism called 'controllable thinking effort.' This capability allows users to adjust computational costs according to processing complexity, enabling them to control the balance between accuracy and cost. While most frontier models maintain proprietary internal architectures and compete on performance, this transparent and controllable design can serve as a differentiation axis.

As another key feature, Thinking Machines explicitly states that its design allows 'direct responses to topics that could be subject to censorship.' The company explains that this could serve as an option for enterprise users prioritizing factual accuracy, particularly in cases demanding reliable outputs regardless of topic sensitivity.

The release of Inkling is set against the backdrop of 2026, when the open-weight model market is rapidly becoming more robust. As a startup led by Mira Murati from OpenAI, the company has garnered attention, and its first model deployment generates significant industry interest. The commercially available license format and design premised on on-premises and private cloud operation expand practical options for enterprises exploring AI implementation in their own environments. Going forward, how well Thinking Machines can maintain the generalist's breadth of coverage at a practical standard will likely remain the focus of ongoing evaluation.