A 27B Open Model Beat Frontier AI at ARC-AGI-3's First Milestone

ARC Prize announced the results of ARC-AGI-3's first milestone prize on July 6, 2026, for a contest period that closed June 30. First place, and $25,000, went to Tufa Labs for a system called the Duck harness, built on Qwen 3.6 27B run locally rather than on any frontier proprietary model. Second place ($10,000) went to a solution called Reki, third ($2,500) to an entrant named Md Boktiar Mahbub Murad with a system called forge. All three were required to open-source their harnesses as a condition of the prize.

ARC-AGI-3 is built to resist the kind of brute-force pattern matching that has eroded earlier ARC benchmarks. Where ARC-AGI-2 posed static puzzles, ARC-AGI-3 drops an agent into a video-game-like environment with no instructions: it has to explore to find out what the rules even are, build a working model of the environment from that exploration, decide on a goal nobody stated, and then plan and act toward it. Kaggle's leaderboard evaluation runs on a single GPU per submission, which caps how much brute compute an entrant can throw at the problem.

That constraint is why the winning approach looked the way it did. The Duck harness has the model write and execute Python in a live REPL, converting the game's state into variables it can reason over directly rather than staring at a rendered image. To keep a long session usable, it manages context by evicting its oldest messages once the conversation grows too large — described by Tufa Labs as playing indefinitely rather than running out of room to think. Perception, when needed, combines the rendered frame with an ASCII grid and a tool for zooming into specific regions.

The two runner-up systems took the more conventional route: a vision-language model reading the rendered board and returning one JSON action per step. Reki's version refreshed a reflection memory roughly every ten steps and used a hardcoded heuristic for exploratory clicks, plus a mechanism the team called a dead signature to stop clicking on object types that had already proven useless. Murad's forge used a similar image-to-JSON pipeline wrapped in a configurable framework with a scoring arbiter choosing between candidate actions — notably, the run that actually won third place had that extra machinery switched off.

On score, the two available figures do not agree, and the difference matters. Tufa Labs put its winning configuration at 1.21 percent; AlphaSignal reported the duck harness scoring 1.03 on the Kaggle leaderboard, which is the number the competition itself ranks on. Either way it is a low figure in absolute terms, and either way it sits above the general-purpose frontier models, which AlphaSignal put under 1 percent on the same benchmark, with GPT-5.4 and Gemini both below 0.5 percent — models that score close to perfectly on the older, static ARC-AGI-1 test. The gap between a 27-billion-parameter open model with a lightweight harness and systems built at far greater scale and cost was, on this measure, essentially closed. It was not opened by much: at 1.03 against 0.5, the lead is a fraction of a percentage point.

Cost is the other axis the milestone was designed to surface. Tufa Labs says its harness ran an order of magnitude cheaper per game than a comparison system it calls Executable World Models, while covering a similar range of games — evidence that on this task, harness design and efficient exploration are doing more work than raw parameter count or inference spend.

The second and final milestone prize for ARC-AGI-3 closes September 30, 2026, with the same $25,000/$10,000/$2,500 structure. Whether a frontier lab enters directly, or the field keeps being won by small open models wrapped in careful harnesses, is what that date will settle.