Inference
EN
Tencent

Hy4 preview

Hy4 preview is a reasoning model from Tencent in the Hy4 family. 27 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2026-08
RankedEnough results for an index.
IndexAll categories, out of 100.66.4±4.1Median 51.0
CoverageIndex weight that has results.90%
SpeedOutput tokens per second.67/sMedian 67/s
First tokenSeconds to the first output token.1.3 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$0.834Median $0.60
Output / 1MUS dollars per 1M output tokens.$2.50Median $2.55
ContextMaximum tokens in one request.1.05MMedian 262K
EloLMArena rating and rank.—Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
tencent/hy4-preview
Max outputMaximum output tokens per request.
64K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
Yes
Cache read / 1MUS dollars per 1M cached tokens.
$0.042
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
3 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Deep InfraHost$0.834$2.501.05M
Novita AIHost$0.834$2.501M
OpenRouterRouter$0.834$2.501.05M

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Run locally

779.96B parameters
16-bit
1638.9 GB
8-bit
871.1 GB
4-bit
497.5 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.16 counted63.552 / 218
CodingCode writing and repair.6 counted65.539 / 209
ReasoningLogic problems and puzzles.1 counted69.6—
MultimodalTasks with images and text.1 counted75.7—
KnowledgeFacts and expert knowledge.2 counted67.929 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.None counted——
MathMath problems.1 counted73.0—

Results

27 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
WideResearchAgentic83.9%1 / 16
Graduate-Level Google-Proof Q&AKnowledge92.3%18 / 141
MCP AtlasAgentic83.7%8 / 52
SWE-bench ProCoding65.7%14 / 83
Humanity's Last ExamKnowledge55.4%14 / 77
Code MigrationCoding47.4% ±4.314 / 77
NL2RepoCoding58.9%7 / 36
Critical Physics TasksReasoning16.9%43 / 214
Terminal-Bench 2.1 · BenchLMAgentic85.4%18 / 86
GPQA Diamond · BenchLMKnowledge92.3%19 / 84
Agent Arena command recoveryAgentic6.512 / 50
JobBenchAgentic61.7%9 / 36
Vibe Code Bench v1.1Coding77.5% ±4.037 / 123
Toolathlon-VerifiedAgentic74.1%8 / 23
ProofBench v1.1Math75.0% ±4.419 / 54
OfficeQA ProMultimodal66.2%6 / 16
Humanity's Last Exam without toolsKnowledge43.4%20 / 53
IOICoding59.3% ±4.617 / 45
Agent Arena task outcomeAgentic3.521 / 54
Agent Arena steerabilityAgentic0.226 / 54
Humanity's Last Exam with toolsAgentic55.4%18 / 36
DeepSWEAgentic64.3%36 / 59
APEX-AgentsAgentic37.1%8 / 13
CWE-bench v2Agentic47.3%7 / 11
Terminal-Bench 2.1 · Vals AIAgentic55.1% ±3.449 / 75
CyberGymAgentic78.4%19 / 29
AutomationBenchAgentic32.1%24 / 32
Terminal-Bench 4.0Agentic8.1% ±1.348 / 61
CWE-bench v1Agentic53.0%10 / 12
Hermes Index: agentic capability and cost per task inside Hermes AgentAgentic26.1%15 / 16
Agents' Last ExamAgentic22.8%21 / 21
ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding17.5%19 / 19
Public Benefits Bench v1.1Public benefits68.6% ±1.29 / 50
Legal Research BenchLegal45.2% ±3.522 / 90
Tax Agent BenchFinance32.6% ±3.421 / 82
Finance Agent (v2)Finance55.1% ±0.323 / 79
LegalBenchLegal83.8% ±0.447 / 149
MedScribeHealthcare83.6% ±2.145 / 109
MedCodeHealthcare43.2% ±2.147 / 106
BioMysteryBenchScience69.3% ±2.615 / 30
EMBFinance58.1% ±3.138 / 75
ProgramBenchCoding0.0%34 / 65
HLAB-Vals-1.1Legal8.8% ±2.223 / 35
CyberBench v1.1Other65.4% ±5.532 / 48
SRE BenchOther2.3% ±0.924 / 36
Terminal-Bench-ScienceScience1.4% ±1.437 / 44
BankerToolBenchAgentic78.6%—
Data Research and Analysis with Complex OperationsAgentic77.2%—
ApexMath74.2%—
Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behindAgentic58.6%—
PostTrain BenchCoding35.6%—
SWE-MarathonCoding31.9%—
STEMBench120: six-domain scientific research progressReasoning2.6%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseVals AINo licence statedLMArenaCC BY 4.0models.devMIT

History

5 releases
Index67.4 → 66.4
23 Oct 202612 Nov 2026
Output / 1M$2.50 → $2.50
23 Oct 202612 Nov 2026

Questions

How much does Hy4 preview cost?
$0.834 per 1M input tokens, $2.50 per 1M output tokens.
How good is Hy4 preview?
Its index is 66.4 out of 100. That is place 35 of 283 ranked models.
What is Hy4 preview best at?
Knowledge: 67.9 out of 100, place 29 of 268.
How fast is Hy4 preview?
67 output tokens per second, 1.3 s to the first token.
What is the context window of Hy4 preview?
1.05M tokens.
Is Hy4 preview open source?
Yes. Its weights are published.
When was Hy4 preview released?
August 2026.

Badge

Hy4 preview on Inference360: place 35 of 283, index 66.4
HTML
<a href="https://inference360.ai/models/hy4-preview"><img src="https://inference360.ai/badges/hy4-preview.svg" alt="Hy4 preview on Inference360: place 35 of 283, index 66.4" height="20"></a>
Markdown
[![Hy4 preview on Inference360: place 35 of 283, index 66.4](https://inference360.ai/badges/hy4-preview.svg)](https://inference360.ai/models/hy4-preview)

Compare with

Inkling 2 Small66.4Sakana Fugu-Ultra66.3Kimi K366.9Ling 3.1 Flash67.2Hy472.0

Lower price

Ling 3.1 Flash67.2 · $0.135/1MMiMo-V2.6-Flash64.3 · $0.28/1MDeepSeek V4.1 Flash 101768.5 · $0.40/1M

More from Tencent

All models
Hy MT2 7BUnrankedAuKUnrankedUI-Mate-27BUnrankedUI-Mate-9BUnrankedHy ASR 3.0 previewUnranked