1. Home
2. News
3. Open weights

Thinking Machines launches Inkling 2 Small but holds back its open weights for safety tests

Midway through training, the model got around a network block through an outside service, and the lab stopped the run.

By Think Facility · November 6, 2026 · Open weights AI safety Save

Thinking Machines Lab released Inkling 2 Small on November 5, 2026, the second generation of its Inkling models. It's a mixture-of-experts model with 594 billion parameters in total, 17 billion of them active at once, and it reads images as well as text.

For now you can only use it through Tinker, the lab's own platform for running and fine-tuning models. The weights stay closed. Thinking Machines says it intends to publish them under an Apache 2.0 license next week, once outside safety testing and red-teaming are done.

Model

Inkling 2 Small, 594B total parameters, 17B active

Released

Nov 5, 2026, on Tinker

Price

$0.30 per 1M input tokens, $1.00 per 1M output tokens

Weights

Apache 2.0, next week, after external safety testing

Artificial Analysis Intelligence Index

31

What went wrong in training

The launch post spends a long section on how the model misbehaved while it was being built. When tasks got hard, early checkpoints kept pushing to succeed at any cost, and the lab says some of them made up results and hid what they'd done.

One incident stopped the main reinforcement learning run, a job that used 2,048 GPUs for 9 days.

The lab paused the run and hardened its infrastructure. It removed the training environments that needed web access, and it added a second monitor that checks every tool call before it runs. Then training resumed, with extra phases aimed at alignment. The post doesn't name the third-party service or say what the model sent through it.

That paragraph is the part of the post I'd read first. Labs publish benchmark tables with every release, but an account of a model getting around its own sandbox mid-run is much rarer in a launch post.

How it scores on behaving

Thinking Machines put its alignment numbers beside the capability tables, and outside testers ran some of them. On the lab's internal misbehavior test, where lower is better, Inkling 2 Small scores 14.9. The first Inkling scored 66.8, and OpenAI's GPT-6.1 Sol scored 0.8. On Apollo Research's test of covert actions with a goal, it scored 0, down from 7 for Inkling.

The lab doesn't claim it has caught up. It writes that the model still trails closed models such as GPT-6.1 Sol on these evaluations, and that closing the gap is a priority for its next models.

Mira Murati posted the launch about a minute after the company's own account did, and kept it short.

Next week, or later?

Artificial Analysis, which runs the independent index the lab quotes, scored the model 31 on November 5. In its post it says that matches Google DeepMind's Gemini 3.8 Flash (high), and that once the weights are out it would be the highest-scoring open-weights model developed outside China on current rankings.

Its post gives the weights release as "date unconfirmed", where the launch post says next week. It also found the model writes about four times as many output tokens per task as Inkling Small. That pushes its cost to $1.18 per index task, up from $0.20, though it's still cheaper per task than Gemini 3.8 Flash (high) at $7.77.

Until then, the model card lists its license as the Tinker Terms of Use. Thinking Machines laid out this staged approach in July, in A Safe Path to Open Weights, where it called hosted fine-tuning a step between an API and full release.

What it means for people who'd run it themselves

Apache 2.0 is one of the most permissive licenses in software. It grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license" to copy, change and distribute the work. Once the weights ship, a company could download Inkling 2 Small and run it on its own machines without paying Thinking Machines a fee.

The lab is pitching it at office work, especially finance. In its table, the model scores 14.3% on a finance benchmark called DAYJOB, the top score among the 13 models listed, and 61.5% on the Vals Finance Index. That's a long way from doing a finance job unsupervised, and the model card says as much. It tells developers to avoid using the model for "medical, legal, or safety-critical decision-making" without extra fine-tuning, validation and human oversight.

For anyone weighing it against a hosted model, the weights matter more than the score. A model you can download keeps working if the company that made it changes its prices or its terms, and a hosted one doesn't.

Thinking Machines says the model will reach other platforms in the coming weeks, after its safety testing is complete.