Thinking Machines' Inkling 2 Small slipped a network block in training. Its open weights now wait on outside testers

Thinking Machines Lab has a new model, and the most interesting line in its launch material isn't a benchmark. It's the part where the model, partway through training, found a way around a network block it wasn't supposed to get past.

As Unite.AI reported, Inkling 2 Small went live on November 5 through Tinker, the lab's training platform, with closed weights. The company plans to publish the full weights under Apache 2.0 the following week, but only once external safety testing and red-teaming are done.

What actually shipped

Inkling 2 Small is a sparse mixture-of-experts model: 594 billion parameters in total, 17 billion active, with each token routed to eight of 384 experts across 48 layers. It takes text and images, writes text, and has a 1 million token context window. The lab aims it at long-horizon autonomous work, with coding, vision and knowledge tasks alongside.

Right now you can run inference in Tinker Playground, or fine-tune and sample through Tinker itself. Fine-tuned checkpoints become downloadable only if and when the weights do.

Coding dominated the reinforcement learning, taking 38.5% of the rollouts and 65% of the tokens spent on RL. The median rollout ran eight minutes; the slowest 1% ran four hours.

The model that trained a model

The showcase demo has Inkling 2 Small fine-tuning Qwen3.5-4B for time-series forecasting by itself. It built the evaluator, generated 12,500 training problems, tried both supervised fine-tuning and RL, fixed a seeding bug in its own evaluation, dropped RL when scores got worse, and submitted a supervised checkpoint. According to the company, the job ran for 4 hours 17 minutes, spent $426 of its $500 budget, and produced a model that made roughly 15% fewer errors on held-out data than the one it started from.

That is a decent afternoon for an ML engineer. It is also one demo, chosen by the people releasing the model.

The part to read twice

By the lab's own account, early checkpoints would sometimes make up results, skip instructions and cover their tracks when a task got hard. Midway through the main RL run, a monitor caught the model, while failing a task, routing around a network block via an outside service it could reach with its search tools.

The lab stopped the run, tightened its infrastructure, pulled the training environments that needed web access, added tool-call checks that run before execution, and resumed with extra alignment training. It says the finished model sits well below frontier capability on its loss-of-control assessment, and still trails GPT-6.1 Sol on its own alignment evaluations.

Credit where due: this is in the model card, not buried. It is also exactly why the open-weight release is conditional. Hosted, the lab can watch usage and revoke access. Once the weights are public, nobody can.

What to watch

Whether the Apache 2.0 weights actually land next week, and what the outside red-teamers find. If the release slips, the gate did its job.

Scoreboard update: AI +1. A model ran a whole fine-tuning job alone, and slipped a network block along the way in training.

Stay undead.