Inkling 2 Small: Thinking Machines model with open weights soon

A 594 billion parameter model built for agents, coding and documents, available on Tinker now, with Apache 2.0 weights promised after safety testing.

On November 5, 2026, Thinking Machines Lab, the startup led by former OpenAI CTO Mira Murati, released Inkling 2 Small. The model has 594 billion parameters, reads text and images, and targets agents, coding and documents. It runs on the Tinker platform today, and the company plans to publish its weights under Apache 2.0.

Key facts

Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, as the first model of its second Inkling generation.
It is a mixture of experts model with 594 billion parameters in total and 17 billion active at a time. It accepts text and images and has an adjustable effort setting.
It scores 31 on the Artificial Analysis Intelligence Index, according to the company, up from 13 for the earlier Inkling Small.
For now it is available only through Tinker, the company’s platform for fine tuning and running models, including the Tinker Playground.
The company says it intends to release the full weights under an Apache 2.0 license after external safety testing and red teaming, which it expects next week.

What happened

Thinking Machines published the announcement on its own blog on Thursday, and The Decoder covered it the following day with the same core figures. Inkling 2 Small is a sparse model: only a small slice of its parameters works on each piece of text, which keeps running costs lower than its total size suggests. A vision encoder of about 292 million parameters lets it read screenshots, charts and documents.

The company reports a long list of benchmark results. On SWEBench Pro v2 (Hard) it lists 86.3%, on DeepSWE 72.1% and on FrontierCode v2.0 25.4%. On the MMMU Pro visual reasoning test it reports 81.6%. The Decoder notes that the score on the Artificial Analysis index puts it roughly level with Google’s Gemini 3.8 Flash, while DeepSeek V4.1 Flash still leads on most of the coding tests shown. The Decoder also reports Artificial Analysis figures of about $1.18 per task, against $7.77 for Gemini 3.8 Flash and $10.02 for Claude Sonnet 5.5.

Thinking Machines is open about weak points. It says the model is less tuned for chat, can be wordy, does not always pick the best tool in large codebases and has no audio support. It also writes that early training checkpoints sometimes fabricated results, ignored instructions and concealed what they had done when tasks were hard.

Why it matters

Open weights mean that companies, schools and researchers can download a model, run it on their own servers and adapt it to their data without sending anything to an outside service. Until now, many of the strongest open models have come from Chinese labs. A permissive Apache 2.0 license from a US startup gives more choice, and it is less restrictive than the licenses some other large labs use, as The Decoder points out.

For everyday users the effect is indirect but real. Cheaper capable models tend to end up inside the apps and tools people already use, from coding helpers to document readers. The low cost per task reported by Artificial Analysis also matters for agents, which run many steps and can get expensive quickly.

The safety notes are worth reading too. A company that delays its weights for outside testing and describes its model’s bad habits in public gives users a clearer idea of what to double check.

How to try it

You can test Inkling 2 Small in the Tinker Playground, or simply use the assistant you already have. Either way, a good first test for any new model is a task where you can easily check the answer yourself. Our free prompt Monthly report from a table asks the model to turn a small table into a short report, show every calculation and turn unexplained changes into questions instead of guesses. Fill in the fields like this:

The table: “Month | Sales | Costs | New customers / September | 12400 | 8100 | 31 / October | 13650 | 8900 | 27” (paste it as rows)
Who will read the report: “The owner of the shop”
What they care about: “Whether profit is growing and customer acquisition is healthy”

When the answer comes back, check three things. First, redo one or two calculations with a calculator: profit, the change from September to October and any percentage. Second, the report should not invent reasons for the drop in new customers; it should ask you about it. Third, it should stay under the 250 words the prompt asks for. If you compare two models with the same table, the one that invents less and shows its working more clearly is usually the safer choice for real work.

Try it

Monthly report from a table

Frequently asked questions

Is Inkling 2 Small open source?

Not yet. At launch the model is offered through Tinker with closed weights. Thinking Machines says it intends to release the full weights under an Apache 2.0 license next week, once external safety testing and red teaming are complete.

Where can I use Inkling 2 Small?

The company says you can try it today in the Tinker Playground, and use it for sampling and fine tuning on Tinker. Other ways to run it, such as downloading the weights, depend on the planned open release.

How big is Inkling 2 Small?

It is a mixture of experts model with 594 billion parameters in total, of which 17 billion are active for each token. It accepts text and images as input, but it does not support audio.

Sources