Hy4 preview

Name: Hy4 preview
Author: Tencent

Tencent

Hy4 preview is Tencent Hy Team's new-generation MoE flagship with 770B total and 49B activated parameters, released under Apache 2.0. It pairs Gated DeepSeek Sparse Attention with IndexCache, identity Hyper-Connections, a native MTP layer for speculative decoding, and a 1,048,576-token context window. A text-only hybrid-thinking model targeting software engineering, office and analysis work, game development, and scientific research, co-designed with Tencent products such as CodeBuddy and WorkBuddy.

Key Specifications

Parameters

770.0B

Context

Release Date

August 28, 2026

Average Score

58.7%

Repository Model Weights Results Blog

Timeline

Key dates in the model's history

Announcement

August 28, 2026

Last Update

August 31, 2026

Today

October 17, 2026

Technical Specifications

Parameters

770.0B

Training Tokens

Knowledge Cutoff

Family

Capabilities

MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Reasoning

Logical reasoning and analysis

GPQA

GPQA Diamond • Self-reported

92.3%

Other Tests

Specialized benchmarks

Agents' Last Exam

ALE-CLI • Self-reported

22.8%

APEX-Agents

pass@1 • Self-reported

37.1%

ArXivMath

• Self-reported

66.6%

AutomationBench

v1.0.6 • Self-reported

32.1%

BankerToolBench

• Self-reported

78.6%

BioMysteryBench

• Self-reported

71.3%

CritPT

Official • Self-reported

16.9%

CyberGym

• Self-reported

78.4%

DeepSWE

• Self-reported

64.3%

DRACO

• Self-reported

77.2%

GDPval-AA

GDPval-AA v2 Elo (max_score 3000) • Self-reported

55.9%

HorizonMath

pass@4 • Self-reported

8.8%

Humanity's Last Exam (no tools, text-only)

No tools, text-only • Self-reported

43.4%

Humanity's Last Exam (with tools, text-only)

With tools, text-only • Self-reported

55.4%

Job Bench

• Self-reported

61.7%

MathArena Apex

MathArena Apex 2025 • Self-reported

74.2%

MCP Atlas

Public set • Self-reported

83.7%

NL2Repo

• Self-reported

58.9%

OfficeQA Pro

• Self-reported

66.2%

OneMillion Bench

With tools • Self-reported

65.4%

PostTrainBench

PostTrainBench V1.1 • Self-reported

35.6%

Program Bench

• Self-reported

17.5%

SkillsBench

Text-only, 79-task subset • Self-reported

62.9%

SuperChem

• Self-reported

66.4%

SWE-bench Multilingual

• Self-reported

82.9%

SWE-Bench Pro

Public set • Self-reported

65.7%

SWE-Marathon

• Self-reported

31.9%

Terminal-Bench 2.1

• Self-reported

85.4%

Toolathlon

Toolathlon-Verified • Self-reported

74.1%

WideSearch

• Self-reported

83.9%

Workspace Bench

• Self-reported

60.2%

License & Metadata

License

apache_2_0

Announcement Date

August 28, 2026

Last Updated

August 31, 2026

Compare Hy4 preview

All comparisons

vs Gemini 3.8 Flash vs MiMo-V2.6-Pro vs Claude Haiku 5.5 vs Claude Fable 5.1 vs Gemini 3.7 Flash vs DeepSeek-V4-Flash-0731 vs Laguna S 2.1 vs DeepSeek-V4.1-Flash

Similar Models

All Models

Hy3

Tencent

295.0B

Best score:0.9 (GPQA)

Released:Jul 2026

Nemotron 3 Ultra (550B A55B)

NVIDIA

550.0B

Best score:0.9 (GPQA)

Released:Jun 2026

Mistral Large 2

Mistral AI

123.0B

Best score:0.9 (HumanEval)

Released:Jul 2024

Price:$2.00/1M tokens

LongCat-Flash-Thinking-2601

Meituan

560.0B

Best score:1.0 (TAU)

Released:Jan 2026

Qwen3 Max Thinking

Alibaba

1.0T

Best score:0.9 (GPQA)

Released:Feb 2026

Price:$1.20/1M tokens

GLM-5.2

Zhipu AI

753.0B

Best score:0.9 (GPQA)

Released:Jun 2026

Price:$1.40/1M tokens

MAI-Thinking-1

Microsoft

1.0T

Best score:0.8 (GPQA)

Released:Jun 2026

Qwen3.5 122B A10B

Alibaba

122.0B

Best score:0.9 (GPQA)

Released:Mar 2026

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.