H3-Image Debuts on AA Image Boards, Leads Open Weights

When H3-Image launched on November 10, it had no entry on either of Artificial Analysis's image leaderboards — and when this topic page went live, all it could say was "no independent benchmarks yet." On November 12–13 it joined both boards: 1066 on text-to-image, 1138 on editing — the top open-weights model on each. It is the first independent third-party score since launch, and the first fact on this topic page that needs rewriting.

Two scores, two ranks

Readings are from the live AA boards as read at 04:07 Beijing time on November 14, 2026. Both variants (Standard and Highspeed) are listed, four points apart and adjacent in rank; on the editing board Standard actually sits 4 points above Highspeed — meaningless inside a ±16 error band.

Best open weights on both boards — AA's own words

This is not our inference; it is the AA boards' own FAQ: on editing, "H3-Image currently leads among open weights models", followed by Fibo 2.0 at 1085 and Qwen-Image-2.1 at 1084. On text-to-image the open-weights runners-up are also Qwen-Image-2.1 (1046) and FLUX 3 Image Klein (1042), with H3-Image more than 20 points clear. On both boards, H3-Image leads the open field — and by more than half an error band.

The gap to the top: still a distance to the first tier

Read sideways, the position is clear: on text-to-image H3-Image trails Seedream 5.5 by 68 points and the leader by almost 200; on editing the gaps shrink to 21 and 72 points. Editing is H3-Image's relatively stronger side — consistent with its positioning: one checkpoint for generation and multi-reference editing, with editing as the headline use case.

How to read these two scores: four cautions

First, Elo is blind human preference, not an objective quality yardstick: voters compare two images and pick the one they like — it measures "which looks better to people", not instruction following, text accuracy or physical correctness. Second, the vote counts are still small: about 3,200 votes on each board, error bands of ±15–16, rank ranges of 21–24 and 8–17; a few thousand more votes could move the rank several places either way. Third, the scores move: the AA boards update continuously, and 1066 and 1138 are readings from the early hours of November 14, not fixed verdicts. Fourth, these are still not capability benchmarks: H3-Image has no GenEval or DPG-Bench score, and the vendor's "under 10 seconds" speed claim has no third-party timing at all — AA doesn't measure time, so the speed question stays open.

Timeline: listed within three days

Launched November 10; on November 11 OrcaRouter's check found no entry on either board; a Medium roundup updated around 08:00 on November 13 already cited AA scores, so the listing happened on November 12–13. Getting a first independent score within three days of launch is itself notable: Seedream 5.5, released November 9, saw its AA scores of 1134 / 1159 appear within its first week too.