Atlas

Models

← All models

Hy3 preview

Tencent · 2026-04-23 · 298.8B parameters

Hy3 preview is Tencent Hy Team's April 2026 preview release of its sparse mixture-of-experts language model, comprising an approximately 295B-parameter backbone and a 3.8B-parameter multi-token-prediction layer, with about 21B active parameters and a 256K-token context length. It supports direct and reasoning modes and was the predecessor to the July 2026 Hy3 release.

Benchmark scores

BenchmarkScore
AA-LCR54.7
AA-Omniscience Index-34.6
Artificial Analysis Intelligence Index33.6
Artificial Analysis Omniscience Accuracy28.0
Artificial Analysis Omniscience Hallucination Rate75.7
Artificial Analysis Openness Index33.3
BullshitBench v2 Leaderboard16.0
BuseyBench SVG2.1
Buyout Game Bradley-Terry Rating1621
CadQueryEval52.0
CadQueryEval — Bounding Box52.0
CadQueryEval — Chamfer Distance52.0
CadQueryEval — Code Execution64.0
CadQueryEval — Component Count64.0
CadQueryEval — Hausdorff 95p52.0
CadQueryEval — STL Creation64.0
CadQueryEval — Volume56.0
CadQueryEval — Watertight Mesh60.0
Capability147.9
CRED v0.9250.5
CRED v0.92 - Data integrity85.0
CRED v0.92 - Execution100.0
CRED v0.92 - Paper–code26.7
CRED v0.92 - Prose–table46.0
CritPt4.6
GBENCH One-shot Coding GScore35.5
GeneBench-Pro0.9
GeneBench-Pro Artificial Analysis Subset0.4
GeneBench-Pro Externally Reviewed0.9
GeneBench-Pro Not Externally Reviewed0.9
Loading Atlas data…