Hy3

Name: Hy3
Author: Tencent

Tencent

Hy3 is a 295B-parameter MoE model with 21B active parameters from Tencent's Hy Team, the successor to Hy3 Preview. Post-training was scaled up with higher-quality data and RL guided by feedback from 50+ products, delivering strong reasoning, agentic, and long-context performance that rivals open models 2-5x its size. Hybrid-thinking with configurable reasoning effort, a 256K context window, and production-grade tool-call and output-format stability.

Key Specifications

Parameters

295.0B

Context

Release Date

July 6, 2026

Average Score

57.1%

Repository Model Weights Results Blog

Timeline

Key dates in the model's history

Announcement

July 6, 2026

Last Update

August 31, 2026

Technical Specifications

Parameters

295.0B

Training Tokens

Knowledge Cutoff

Family

Capabilities

MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Programming

Programming skills tests

SWE-Bench Verified

• Self-reported

78.0%

Reasoning

Logical reasoning and analysis

GPQA

GPQA Diamond • Self-reported

90.4%

Other Tests

Specialized benchmarks

AA-LCR

• Self-reported

73.4%

APEX-Agents

pass@1 • Self-reported

25.6%

ArXivMath

• Self-reported

52.2%

BrowseComp

• Self-reported

84.2%

Claw-Eval

pass^3 • Self-reported

68.5%

CL-bench

• Self-reported

23.8%

CL-bench (Life)

Life • Self-reported

17.0%

CMT-Benchmark

• Self-reported

37.9%

DeepSearchQA

• Self-reported

91.0%

DeepSWE

• Self-reported

28.0%

FrontierScience Olympiad

• Self-reported

74.8%

FrontierScience Research

• Self-reported

21.3%

HorizonMath

pass@12 • Self-reported

7.1%

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

No tools, text-only • Self-reported

47.0%

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

With tools, text-only • Self-reported

53.2%

IMO-AnswerBench

• Self-reported

90.0%

MathArena Apex

• Self-reported

38.7%

MCP Atlas

Public Set • Self-reported

79.1%

NL2Repo

• Self-reported

45.6%

PHYBench

• Self-reported

77.4%

SkillsBench

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

55.3%

SuperChem

• Self-reported

54.9%

SWE-bench Multilingual

• Self-reported

75.8%

SWE-Bench Pro

• Self-reported

57.9%

Terminal-Bench 2.1

• Self-reported

71.7%

Toolathlon

• Self-reported

48.5%

USAMO 2026

• Self-reported

72.0%

WideSearch

• Self-reported

76.4%

WildClawBench

Text-only, 35-task public set • Self-reported

53.6%

License & Metadata

License

apache_2_0

Announcement Date

July 6, 2026

Last Updated

August 31, 2026

Compare Hy3

All comparisons

vs Hy4 preview vs Laguna S 2.1 vs IBM Granite 4.2 8B vs Claude Opus 5 vs GLM-5.3 Flash vs GPT-5.6 Terra vs Inkling-Small vs Claude Sonnet 5

Similar Models

All Models

Hy4 preview

Tencent

770.0B

Best score:0.9 (GPQA)

Released:Aug 2026

LongCat-Flash-Thinking

Meituan

560.0B

Best score:0.8 (TAU)

Released:Sep 2025

LongCat-Flash-Thinking-2601

Meituan

560.0B

Best score:1.0 (TAU)

Released:Jan 2026

MAI-Thinking-1

Microsoft

1.0T

Best score:0.8 (GPQA)

Released:Jun 2026

Nemotron 3 Ultra (550B A55B)

NVIDIA

550.0B

Best score:0.9 (GPQA)

Released:Jun 2026

DeepSeek-V4-Flash-0423

DeepSeek

284.0B

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$0.14/1M tokens

Mistral Large 2

Mistral AI

123.0B

Best score:0.9 (HumanEval)

Released:Jul 2024

Price:$2.00/1M tokens

LongCat-Flash-Chat

Meituan

560.0B

Best score:0.9 (MMLU)

Released:Aug 2025

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.