Farang is advancing FX toward superintelligence. The next phase will combine broader world knowledge with the architecture’s reasoning capabilities, building on results across forecasting, vision and structured problems.
One architecture across modalities
FX is modality-agnostic by design. Its architecture is not tied to language, images or any single type of input.
FX2 now ranks first on TS Arena’s live time-series forecasting leaderboard. The same model now achieves results including:
- 90.1% top-1 accuracy on ImageNet-1K strict — visual recognition.
- 99.8% solve rate on Sudoku-Extreme — constraint solving.
- 100% solve rate on Maze-Hard — spatial reasoning.
- 69% Stockfish agreement on Epoch AI’s Chess NBM — next-best-move selection.
- #1 on TS Arena - time series forecasting.
These results include wins over both specialized and general-purpose models. Their breadth is why we are scaling FX rather than limiting it to individual applications. The same architectural approach has proved effective across substantially different problems.
Beyond predicting outcomes
Working across modalities is part of the objective. Another is moving from recognizing patterns and predicting outcomes to finding the actions needed to reach a goal.
This is an important distinction from predictive world models. A predictive world model answers “What happens if I take this action?” It models the consequences of an action supplied to it.
That does not, by itself, answer “Which actions will get me to this goal?” A planner or learned policy must supply that decision-making. An accurate prediction of an environment’s behavior is not itself a plan.
FX integrates prediction with learned action selection. It learns to find intermediate steps toward a desired outcome and predict their consequences. The reasoning takes place internally, rather than through a sequence of generated text.
Our objective is to develop that ability across domains, not just within a particular environment.
Next phase
We are expanding FX’s world knowledge through larger-scale language modeling and testing its reasoning on more complex tasks. The priority is to extend its capabilities while retaining its inference efficiency.
Superintelligence is the objective guiding this work. The architecture and training methods remain proprietary; selected results will continue to be published.