Atlas

Models

← All models

Hy3

Tencent · 2026-07-06 · 298.8B parameters

Hy3 is Tencent Hy Team's July 2026 open-weight mixture-of-experts language model with an approximately 295B-parameter backbone, a 3.8B-parameter multi-token-prediction layer, about 21B active parameters, and a 256K-token context length. It follows the Hy3 preview release with scaled post-training and RL, targeting stronger reasoning, agentic coding, search, long-context, and production tool-use reliability.

Benchmark scores

BenchmarkScore
AA-LCR66.7
AA-Omniscience Index-18.5
AHK-Eval61.1
AHK-Eval Algorithms66.7
AHK-Eval Data Structures50.0
AHK-Eval Date and Time66.7
AHK-Eval Easy83.3
AHK-Eval Hard33.3
AHK-Eval Hidden Cases61.9
AHK-Eval Mid75.0
AHK-Eval Numbers83.3
AHK-Eval Parse Success88.9
AHK-Eval Regex66.7
AHK-Eval Strings66.7
ALE-Bench1516
Arena Code WebDev — Brand & Marketing1526
Arena Code WebDev — Consumer Product1477
Arena Code WebDev — Content Creation Tools1516
Arena Code WebDev — Data & Analytics1482
Arena Code WebDev — Gaming1545
Arena Code WebDev — HTML1429
Arena Code WebDev — React1533
Arena Code WebDev — Reference-Based Design1517
Arena Code WebDev — Simulations1532
Arena Text — Business, Management & Financial Ops (No Style Control)1427
Arena Text — Chinese (No Style Control)1508
Arena Text — Creative Writing (No Style Control)1407
Arena Text — English (No Style Control)1451
Arena Text — Entertainment, Sports & Media (No Style Control)1401
Arena Text — Exclude Ties (No Style Control)1437
Loading Atlas data…