Where the models you can actually use sit, on two indexes that publish their methods.
Open weights only · Artificial Analysis index · size in grey
ARC-AGI-2 and ARC-AGI-3, percent solved · frontier models at max effort
Open weights, 4B to 40B · Progress Index against size · up and to the left is better
Frontier: K2 Horizon 7B (7B, 263) → Qwen3.8 27B (27B, 303) → Nanbeige4.3 (33B, 306). Everything else in that range scores the same or worse at a larger size.
These are independent scores for models anyone can use today. None of them are ours, and none of them are a result for Tachyon. Our own pilot measurements stay in the report until they can be published with their conditions attached.
Source: artificialanalysis.ai — Intelligence Index v5.0.1, Progress Index and parameter counts are Artificial Analysis; ARC-AGI-2 and ARC-AGI-3 are the ARC Prize leaderboard. Retrieved 10 November 2026.