← Slices

Thinking Machines Releases Inkling 2 Small Weights Under Apache 2.0

Thinking Machines Releases Inkling 2 Small Weights Under Apache 2.0

Key takeaways

  • Thinking Machines Lab posted the Inkling 2 Small weights to Hugging Face on November 12, 2026, under an Apache 2.0 license.
  • The 594-billion-parameter mixture-of-experts model activates about 17 billion parameters per token and accepts text and image input.
  • At four-bit precision the weights alone need roughly 297 GB, which puts self-hosting out of reach for most solo operators.
  • An Apache 2.0 checkpoint is a dependency you control, because the vendor cannot quietly reprice or withdraw the copy you already downloaded.
  • For most one-person shops, renting inference from an API provider remains the sensible starting point, with smaller models in the family as the cheap first test.

What happened, and when

On November 5, 2026, Thinking Machines Lab launched Inkling 2 Small through its Tinker platform with closed weights, promising to publish the full model under Apache 2.0 once external safety testing and red-teaming were finished. The weights reached Hugging Face on November 12, a date confirmed by CASRAI and AlphaSignal.

The specs, as of the November 12 release

Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters, roughly 17 billion of them active at a time, with image input and a 1 million token context window, according to AlphaSignal. On the Artificial Analysis Intelligence Index it scores 31, up from 13 for the original Inkling Small. Those launch-era figures come from the lab and its own test setups, so treat them as a starting point rather than a verdict.

Licensing is the headline for operators. Apache 2.0 permits commercial use without a license fee, and the release is spreading through hosting partners: Superpower Daily reports that Together AI already lists the model under Apache 2.0.

The hardware math nobody wants to do

Run the numbers. A 594-billion-parameter model stored at four bits needs roughly 297 GB for the weights alone, before the runtime buffers and key-value cache that long contexts demand. That is a rack-of-GPUs problem, not a laptop problem. It is also why, for most of us, the hosted price matters more than the license.

Local models are getting easier to run. The hardware bill still sets the limit.

Where it leads, and where it does not

Coverage is split on the scorecard. AlphaSignal reports that the model leads its class on several vision benchmarks and trails DeepSeek on coding. A separate developer writeup argues it is not the leader for terminal-based coding-agent work, which is the job many solo builders care about most. Both can be true at once, which is exactly why you should run your own evals before a production workflow depends on it.

What the safety section says

According to CASRAI's reading of the model card, the safety evaluations covered sycophancy, manipulation, refusals, CBRN, cyber and loss-of-control risks, and concluded the model offers no material uplift beyond what the open-weight ecosystem already provides. That is the lab's own evaluation, so read the card yourself rather than taking a summary's word for it.

What it means for solo operators

Three practical points. First, this is a new bargaining chip: a US-developed, Apache-licensed model you can move between hosts without rewriting your stack. Second, a pinned checkpoint is a dependency you control, which matters the day a hosted vendor changes its pricing or terms. Third, self-hosting at this scale is a decision for a company with a hardware budget, not a weekend project. The smaller models in the same family are where most one-person shops should start, and our guide to when the cheap model is the right call explains how to decide.

Local AI, in context

Inkling 2 Small is one more data point in a longer trend. The tooling for running open weights on ordinary hardware keeps improving, from the llama.cpp v1.0.0 milestone to Transformers support for llama.cpp quants. Running a model locally also changes your data exposure, which is why the network egress story is worth reading before you wire any local model into sensitive data.

Our take

After more than 25 years in digital marketing, I have read enough "open" announcements to check the fine print twice. Apache 2.0 is the good kind of fine print. It means the model survives the day its vendor changes its mind, which is more than most of us can say about the SaaS tools we already depend on. But a 594-billion-parameter download is not a business strategy for a one-person shop, any more than a free puppy is a childcare plan: adorable, technically free, and immediately hungry for resources.

My prediction: most solo operators will consume this model through an API for the next year, then discover that a smaller sibling handles the routine work at a fraction of the bill and the hardware. Use the big model for the hard reasoning you cannot yet hand off, and keep the cheap one on the boring jobs. Benchmark both on your own tasks, not the lab's, and keep the receipts. The open license is a gift. Treat it like one, and read the card.

Sources

Frequently asked questions

When were the Inkling 2 Small weights released?

Thinking Machines posted the weights to Hugging Face on November 12, 2026, under an Apache 2.0 license. The model had launched a week earlier on hosted access only, on November 5.

Can a solo operator run Inkling 2 Small on a laptop?

Not realistically. At four-bit precision the weights alone are roughly 297 GB before runtime overhead, so most operators will reach the model through a hosted provider.

What license does Inkling 2 Small use?

The weights are released under Apache 2.0, which permits commercial use without a license fee. Hosting providers such as Together AI list the model under the same terms.

How does Inkling 2 Small benchmark?

It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the original Inkling Small. AlphaSignal reports it trails DeepSeek on coding, so test it on your own tasks.

Related reading

Book a 30-min discovery →