A Soon-to-Be-Free SI Just Doubled Its Score on a Puzzle Test Built to Stump Machines.

A new SI model from Thinking Machines Lab just posted the best open-weight score ARC Prize has ever measured on its hardest pattern test.

On Nov. 6, ARC Prize verified that Inkling 2 Small scored 81.7% on ARC-AGI-2 and 96.5% on ARC-AGI-1 at its highest reasoning setting. The lab charges about 27 cents per ARC-AGI-2 task and 8 cents per ARC-AGI-1 task. ARC Prize says those marks beat every open-weight model it has tested on both exams.

What the test actually is

ARC-AGI is a puzzle exam built to check whether a system can spot a new pattern and apply it, not just recall something it has seen before. Each task shows a few colored grids that follow a hidden rule. The model has to figure out the rule and finish a new grid. Humans find most of these easy. Older SI models often fail them.

ARC-AGI-2 is the harder sequel. A high score there is a real signal that the model can generalize, not just grind memorized answers.

The scoreboard

Inkling 2 Small’s 81.7% tops the open-weight field ARC Prize lists today. DeepSeek’s V4.1 Flash 1017 sits next at 78.1%. Meta’s Muse Spark Open is at 57.5%. Thinking Machines’ own July model, Inkling Small, managed just 40.1%.

That is more than a double from the same lab’s July release. ARC Prize put the jump at 40.14% to 81.67% on the semi-private set.

It still trails the closed giants

Honest ceiling: this is an open-weight record, not the overall record. Closed models on the same ARC Prize board still sit higher. OpenAI’s GPT-6 family shows 95.0% on ARC-AGI-2. Google’s Gemini 4 Argon shows 91.3%. Anthropic’s Claude Opus 5.5 shows 93.3%. Inkling 2 Small closed a lot of the open-weight gap. It has not erased the closed-model lead.

What shipped, and what hasn’t

Thinking Machines introduced Inkling 2 Small on Nov. 5. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token. It takes text and images, supports adjustable reasoning effort, and is live now on the lab’s Tinker platform for inference and fine-tuning.

The weights are still closed at launch. The lab says it plans a full Apache 2.0 release after more outside safety testing. ARC Prize’s own post said “next week.” The lab’s product page softens that to “the coming weeks.” Treat both as a plan, not a delivery date.

What it can actually do

Beyond the puzzle scores, the lab’s own demos show the model building interactive software from scratch and checking its own screens as it works. One demo grows a living city simulator from a single road. Another turns a hand-drawn paper valley into a 3D scene with origami cranes. Those are marketing demos, not the ARC test, but they show the same pitch: long-horizon work with vision in the loop.

The model card also lists strong marks on coding, finance and visual-creation benches the lab reports. Those are the lab’s own tables. The ARC-AGI numbers above are the ones ARC Prize ran itself.

The fine print

ARC Prize has not finished ARC-AGI-3 testing for this model. Weights are not downloadable yet, so outside labs cannot re-run every number from scratch until that drop lands. And a high ARC score is one kind of intelligence test — it does not guarantee the same lead on your coding, chat or agent job.

Still, for anyone watching open models, the verified jump is the story: 40% to 82% on the hard exam, cheapest verified open-weight lead ARC Prize has posted, with the weights promised soon.