Murati is open-sourcing her 594-billion-parameter model

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has announced Inkling 2 Small, its second-generation multimodal model. The company said it will release the weights of the 594-billion-parameter model to the public under the Apache 2.0 license after security testing. For now, the model has been made available to developers for testing through the enterprise platform Tinker.

Huge parameter pool, small active load

Despite the “Small” in its name, Inkling 2 Small has a total of 594 billion parameters. Its cost advantage comes from a sparse Mixture of Experts architecture. The structure includes 384 subnetworks specialized in areas such as math, code, logic, and visual parsing. As each token is processed, the routing layer activates only eight of them; in other words, only about 17 billion parameters are engaged during inference.

The model has a context window of 1 million tokens. That means hundreds of pages of technical documentation, entire code repositories, or long video streams can be kept within a single context block. The company says this design also makes it economical to host the model on enterprise servers.

31 points in independent testing

Inkling 2 Small scored 31 points on the independent intelligence index of AI analysis platform Artificial Analysis. That result doubles the score of the first-generation Inkling models. In evaluations, the model stands out as the first major Western model in the open-weight category able to compete in the same league as Chinese models like DeepSeek and Qwen.

Among the areas where the model stands out are native visual reasoning that directly understands visual inputs, autonomous coding, long-document analysis, and agent workflows capable of managing external software tools. The company says the margin of error on the tool-calling side is very low.

Murati’s open-source choice

In the announcement released alongside the launch, Murati argued that frontier-level models should not remain the monopoly of just a few trillion-dollar tech giants. Murati said:

“The path to artificial general intelligence cannot be locked behind closed API walls. By opening Inkling 2 Small’s weights to the community, we are making sure independent developers, universities, and engineers around the world can keep this power in their own hands.”

The Apache 2.0 license means commercial companies can train the model on their own private data and integrate it into their products without royalties or additional restrictions. The weights are expected to be downloadable via Hugging Face next week.

Why it matters

Over the past two years, open-weight models have largely seemed to be in the hands of Chinese labs. Releasing a 594-billion-parameter model under Apache 2.0 means European and American teams can also run a foundation model at the same scale on their own infrastructure. That expands institutions’ options in terms of both cost and data control.

On the other hand, making the model’s weights publicly available also brings misuse and oversight debates with it. For that reason, Thinking Machines Lab has made it a requirement to complete independent security and red-team testing before release.

Source: WebTuring