Short answer: on 5 November, Thinking Machines said it would release the full weights of a new model, Inkling 2 Small, under the Apache 2.0 licence "next week". That is only a promise so far. If it lands, it gives Indian businesses another capable model they can host themselves or buy from several providers. It is unlikely to change what a 30-person company runs in its own office.
On the morning of 9 November, Thinking Machines' Hugging Face page did not list them, so treat everything below as conditional. Its earlier Inkling-Small did come out under Apache 2.0, but that does not guarantee this one will.
What was actually announced
Thinking Machines' own post describes a Mixture-of-Experts model, a design that uses only part of the model at each step, with 594 billion parameters in total and 17 billion active. It takes images and text as input and does not support audio.
Today the weights are closed and the model is available through the company's Tinker platform. The company says the full weights follow "external safety testing and red-teaming", and that other platforms will carry the model in the coming weeks after that testing. The release is tied to safety work, so the date can move.
What it means in practice
It is not a laptop model. The company gives no hardware requirements, so here is a rough guide. Inkling-Small, at 266 billion parameters, is about a 532 GB download, or about 171 GB in a compressed version. This model has more than twice the parameters, and the 17 billion active figure doesn't shrink what has to be stored, so our own estimate is several hundred GB at the least. Meta has shown a larger open model running on two top-end Mac Studios, but those machines cost far more than a normal office PC. For most firms, this is a job for a hosting provider. Our earlier piece on running AI on your own machines covers that demo and the 7 to 14 billion parameter class that does fit office hardware.
"Open" widens your choices, not your hardware. Once weights are public, other companies can host the model, so you can compare them on price and server location. For an owner, the practical gain is a choice of provider and an exit if one raises prices or changes limits. That matters because the bundled AI inside office software now comes with usage caps, as we covered in the limits piece.
The benchmark tables are the company's own, and they are mixed. Among the five open models in the post's tables, Inkling 2 Small leads on some tests, such as JobBench (70.5%) and the Vals Legal Index (73.9%). DeepSeek-V4.1-Flash beats it on one coding test (77.0% against 72.1%) and a knowledge-work test (1746 against 1630), and Kimi K3 beats it on another coding test (88.2% against 86.3%). On alignment evaluations, which test how a model behaves, the post says it beats comparable open models on the company's internal tests but still trails closed models such as GPT-6.1 Sol. Read that as "competitive", not "the new best". The company also says the model is less optimised for chat, so test it before you put it in front of staff or customers.
What Apache 2.0 gives you
If the release goes ahead under Apache 2.0, the licence is permissive. It grants a perpetual, worldwide, royalty-free copyright licence, allows modification and redistribution, and includes a patent grant. Its text has no acceptable-use restrictions. For now, though, the model card lists the licence as the Tinker Terms of Use and also links a Model Acceptable Use Policy, which sets conditions on using the company's model weights. Read both when the weights come out.
The conditions are small: keep the licence and attribution notices, and mark files you changed. The work comes "AS IS", with no warranty, so if the model gives a wrong answer to your customer, that risk stays with you and your contract with whoever hosts it. The licence page doesn't address training data.
A five-point check before you build on it
- Read the licence on the model's official page once the weights are out, and skip the blog summaries. See whether the Model Acceptable Use Policy still applies, since some releases add their own terms.
- Ask the hosting provider where your data goes: which country, whether prompts are stored, and what the contract says about reuse. An open licence does nothing for you if a middleman reads your files.
- Test it on 20 of your own documents (invoices, contracts, enquiry emails) and compare it with what you use today. Benchmarks measure other people's tasks.
- Price a full month of your real volume, and don't judge from a demo. Ask each provider for its per-token price.
- Keep the model replaceable. Put the model name in one setting, keep your prompts and checks in your own files, so changing models means a new setting and a re-test.
Where hype and usefulness part ways
The headline will be that anyone can download it. The useful part for a business owner is quieter: another credible option, a lower chance of being locked to one vendor, and a reason to ask your current supplier what they would do if you moved. Whether Inkling 2 Small itself is the right choice for your invoices or sales follow-ups is something only a test on your own work can tell you.
Arthat AI
Working on something similar at your company?
Arthat AI builds these automations for Indian businesses every week. If you want to talk through whether this fits your situation, we'll give you a straight answer — even if the answer is "don't build it."
