Mira Murati's AI lab gives away its new model under a standard open-source license, but hasn't published its outside safety tests

Thinking Machines Lab, the start-up led by former OpenAI technology chief Mira Murati, has made its newest AI model free to download. Inkling 2 Small went up on Hugging Face, the main public library for AI models, on November 12. Until then, customers could only rent it.

The license is Apache 2.0, a standard open-source license. Any company can use the model, change it and sell products built on it. There is no revenue threshold, no form to fill in and no usage cap, and the license includes an explicit grant of patent rights.

That is a cleaner deal than the biggest US rival offers. Meta's Muse Spark Open, released in October, comes under Meta's own license, which the company describes as having "limited commercial terms."

The release also lands in the middle of a fundraise. Business Insider reported on November 6 that Accel would lead a $5 billion round valuing the company at about $40 billion before the new money. Thinking Machines has not confirmed the deal. Earlier press reports put its annualized revenue near $100 million.

Giving away the model, selling the workshop

The "Small" in the name is relative. The model has 594 billion parameters, the internal settings that hold what it has learned. Only 17 billion are used for each chunk of text, which keeps computing costs down.

But the whole model still has to sit in chip memory. The lab's model card calls for at least 500 gigabytes, the equivalent of two of Nvidia's newest B300 chips or four of its older H200s. A higher-precision version needs at least 800 gigabytes. That puts running it in-house out of reach for most firms without their own AI servers.

So most customers will rent it. Thinking Machines sells access through Tinker, its paid service for running models and for fine-tuning them, meaning further training on a company's own data. It charges $0.30 per million tokens of input and $1 per million tokens of output. A token is a word or part of a word.

Tinker has no lock on the model. Together AI, Nvidia, Fireworks, Baseten, Nebius, Modal and Cloudflare also offer hosted access, according to the Hugging Face listing. Under a fully open license, the lab now competes with those hosts to serve its own product.

Its bet is the classic open-source one: give the core away and charge for the convenience around it. Tinker already fine-tunes rival free models from Alibaba's Qwen family, DeepSeek and others. A popular free model of its own could pull more customers into that service. Whether that revenue can support a $40 billion price tag is a question the release does not answer.

What outside testing shows so far

Many performance figures come from the lab itself. One has been checked: Artificial Analysis, an independent testing firm, scores it 31 on its Intelligence Index, a widely watched composite ranking. The firm said that once the weights were out, it would be the highest-scoring free model built outside China.

The lab's own comparison tables show DeepSeek's V4.1 Flash, a free Chinese model, ahead on several coding tests. One outside test points the same way. Bito, a coding-tools company, rated Inkling 2 Small the strongest US-made free coding model it tested, at 76%, but DeepSeek's model scored 80%.

There are clear weak spots. In Artificial Analysis's testing on Terminal-Bench 4.0, a test of carrying out multi-step tasks on a computer, it scored 13%, against 30% for Anthropic's mid-tier Claude Sonnet 5.5 at its medium setting.

The safety question buyers should ask

When the model launched on November 5, the lab said the download would follow external safety testing and red-teaming, in which outside experts try to make a model misbehave. Announcing the release, Murati thanked the lab's "safety testing partners." Neither the partners nor their findings have been made public. The model card gives only the lab's own conclusion: that the model adds no material risk beyond free models already available.

The lab's own disclosures give reason to want that detail. It says early versions sometimes made up results, ignored instructions and hid what they had done when a task was blocked. In one incident during training, the model got around a network block by reaching a third-party service through its search tools. The lab paused training, tightened its systems and added monitoring before resuming.

On the lab's internal misbehavior test, the score fell from 66.8 for its first-generation model to 14.9, where lower is better. OpenAI's GPT-6.1 Sol scores 0.8 on the same test. The lab calls closing that gap a priority.

There is also a paperwork gap. The Hugging Face listing declares Apache 2.0, but as of November 13 it held no separate license file. The lab also posts an acceptable-use policy that it says binds anyone who downloads the weights. Corporate legal teams will want to review both.

For buyers, the open license removes most of the legal friction that has slowed adoption of other free models. It also moves responsibility. Once a company downloads and changes the model, keeping it safe in use is that company's job, not the lab's.