Skip to main content

Mira Murati Strikes Back: Think Machine Lab Unveils Inkling, a Powerful Open-Source AI

Mira Murati, the former chief technology officer of OpenAI, has officially re-entered the AI arena with a splash. Her newly founded startup, Think Machine Lab, just released its first major creation: Inkling, a multimodal open-source model that's already being hailed as one of the strongest competitors in the US open-source AI scene.

A Team of Heavy Hitters

The lab's founding team reads like a who's who of AI talent. About two-thirds of the core members come straight from OpenAI, where they worked on cutting-edge research, product development, and safety mechanisms. It's a group that knows its stuff—and it shows.

Inkling: By the Numbers

Inkling is built on a mixture-of-experts (MoE) architecture, boasting a total of 975 billion parameters. But here's the kicker: only 41 billion are activated per task, making it efficient despite its size. It also supports a context window of up to 1 million tokens—enough to process entire books in one go.

During pre-training, the model devoured an enormous dataset spanning text, images, audio, and video. Think Machine Lab has now released the full model weights on Hugging Face and through its own Thinker API, so developers can dive right in.

Benchmark Buster

In head-to-head tests, Inkling has proven its mettle across multiple domains. On the AIME 2026 math test, it scored a stunning 97.1%, edging out DeepSeek V4Pro. In the MCP Atlas benchmark for agent workflows, it crushed Nemotron3Ultra with a score of 74.1%.

Where does it fall short? In reasoning, coding, and specialized software engineering tasks, closed-source giants like Claude Fable5 and GPT-5.6Sol still hold an edge. But Inkling shines in native visual and audio understanding. On the MMAU audio benchmark, it scored 77.2%, coming very close to Gemini 3.1 Pro.

Open-Source Showdown

In the open-source arena, Inkling faces tough rivals like GLM-5.2, DeepSeek V4Pro, and Kimi K2.6. GLM-5.2, for instance, still leads in pure coding and complex reasoning. But Inkling pulls ahead in general instruction following (IFBench), showing its versatility.

What This Means

For the open-source community, Inkling is a gift. It provides a powerful base model with native multimodal capabilities—something that's been rare in the open-source world. With its weights freely available, developers can fine-tune it for everything from chatbots to video analysis.

Murati's comeback is more than just a personal victory. It signals that the open-source AI movement is alive and kicking, with serious talent and resources behind it. Inkling may not be the absolute best at everything, but it's a formidable all-rounder that's sure to accelerate innovation.

Key Points

  • Founder: Mira Murati, former OpenAI CTO, leads Think Machine Lab.
  • Model: Inkling is a multimodal open-source AI with 975B total parameters (41B activated).
  • Performance: Top scores in math (97.1% on AIME 2026) and audio (77.2% on MMAU).
  • Availability: Model weights released on Hugging Face and Thinker API.
  • Competition: Strong against open-source rivals, but trails closed-source leaders in some tasks.

Image