Venice Atlas
← Back to model catalog

VENICE ATLAS · MODEL DOSSIER · HY4 PREVIEW

Model overview

Hy4 Preview

Hy4 Preview is a 770B-parameter Mixture-of-Experts flagship model from the Tencent Hy team, with 49B parameters active per token and a 1M-token context window. It is post-trained for software engineering, office and data analysis, game development, and scientific research, with strong long-horizon reasoning, tool use, and agentic performance.

hy4-preview · venice.ai
Text · Venice catalog$1.04 per 1M tokensPrice on Venice AI
ModalityTextOn Venice AI
Primary price$1.04 per 1M tokensper 1M tokens
Context window1.0MMax tokens
Benchmarks32Benchmark results

Benchmark results

See the tests behind each score. They use the benchmark’s original scoring scale.

Agents#1934.80 ratingBiology#2730.86 ratingChemistry#638.74 ratingCoding#1836.99 ratingGeneral intelligence#1849.57 ratingKnowledge#715.49 ratingLong Context#2218.15 ratingMath#1738.50 ratingPhysics#2034.85 ratingProductivity#414.02 ratingReasoning#1449.69 ratingSafety#169.08 ratingScience#139.90 ratingSearch#827.49 ratingSystems#88.36 ratingTool Calling#1428.72 rating
Agents11 results
Agents' Last Examagents-last-exam · agents
22.8%
ReportedOct 2026
AutomationBenchautomationbench · agents
32.1%
ReportedOct 2026
BankerToolBenchbankertoolbench · agents
78.6%
ReportedOct 2026
DRACOdraco · agents
77.2%
ReportedOct 2026
Job Benchjob-bench · agents
61.7%
ReportedOct 2026
OfficeQA Proofficeqa-pro · agents
66.2%
ReportedOct 2026
OneMillion Benchonemillion-bench · agents
65.4%
ReportedOct 2026
PostTrainBenchposttrainbench · agents
35.6%
ReportedOct 2026
Program Benchprogram-bench · agents
17.5%
ReportedOct 2026
Toolathlontoolathlon · agents
74.1%
ReportedOct 2026
Workspace Benchworkspace-bench · agents
60.2%
ReportedOct 2026
Biology1 result
BioMysteryBenchbiomysterybench · biology
71.3%
ReportedOct 2026
Chemistry1 result
SuperChemsuperchem · chemistry
66.4%
ReportedOct 2026
Coding5 results
CyberGymcybergym · code
78.4%
ReportedOct 2026
DeepSWEdeepswe · code
64.3%
ReportedOct 2026
NL2Reponl2repo · code
58.9%
ReportedOct 2026
SkillsBenchskillsbench · code
62.9%
ReportedOct 2026
SWE-Marathonswe-marathon · code
31.9%
ReportedOct 2026
General intelligence1 result
GDPval-AAgdpval-aa · general
55.9%
ReportedOct 2026
Math3 results
ArXivMatharxivmath · math
66.6%
ReportedOct 2026
HorizonMathhorizonmath · math
8.80%
ReportedOct 2026
MathArena Apexmatharena-apex · math
74.2%
ReportedOct 2026
Reasoning8 results
APEX-Agentsapex-agents · reasoning
37.1%
ReportedOct 2026
CritPTcritpt · reasoning
16.9%
ReportedOct 2026
GPQAgpqa · reasoning
92.3%
ReportedOct 2026
Humanity's Last Exam (no tools, text-only)humanity's-last-exam-(no-tools,-text-only) · reasoning
43.4%
ReportedOct 2026
Humanity's Last Exam (with tools, text-only)humanity's-last-exam-(with-tools,-text-only) · reasoning
55.4%
ReportedOct 2026
MCP Atlasmcp-atlas · reasoning
83.7%
ReportedOct 2026
SWE-bench Multilingualswe-bench-multilingual · reasoning
82.9%
ReportedOct 2026
SWE-Bench Proswe-bench-pro · reasoning
65.7%
ReportedOct 2026

Learn what each benchmark measures ↗