Release Date 2026-11-041.0M IN · 64K OUTOpen Model(Apache 2.0)API available

Hy4 is Tencent's latest large language model, the full release that follows the earlier preview version. The biggest change is that it now understands text and images together; the company says this extends to fine-grained image processing, chart reading and even operating programs on its own by looking at the screen. The company also says it follows instructions more precisely, reads long documents in one go and carries out multi-step tasks on its own more reliably than before. In AIB's use-case ratings it sits in the upper tier for documents, while other uses do not yet have enough comparable tests to name strengths or weaknesses. Because most of the results and examples come from the company itself, it is wise to try the model on your own material before relying on it for real work.

API cost

Market price (median)
$0.834IN$2.5OUT
Price range5 providers
$0.751~$1.04IN$2.25~$3.13OUT
SourceOpenRouterLast updated

Real-world usage

Outside weekly top 50
3.1T8/24
15T8/31
17T9/7
13T9/14
9.6T9/21
6.5T9/28
8.3T10/5
7.6T10/12
9.0T10/19
9.1T10/26
3.5T11/2

Output Speed

Slowper 2,000 tokens
53tok/s
First Output 1.6stotal: 39.3s
SourceOpenRouterLast updated

User Comparisons

Writing Style
Answer length
Very long
top 17%
Vocabulary diversity
Plainer side
bottom 37%
Sentence rhythm
Varied side
top 39%
Language mixing
High
top 22%
List usage
Often
25% of lines
Heading usage
Occasionally
12% of lines
Reasoning share
Very heavy
59% of total
Response Speed
instability ±203%
First response 32.2sTotal 47.8s
116 tok/s
6.9ms per token
SourceAIB Inc.Last updated

Expert Performance Evaluation

How high is its intelligence index across knowledge work, coding and science?AA Intelligence Index v5Top 43%#23
34.6

More about AA Intelligence Index v5

Can it pick the right tools and use them properly?MCP-AtlasTop 17%#6
83.7%Reasoning effort: MaxSelf-reported by developer

More about MCP-Atlas

61.7%Reasoning effort: MaxSelf-reported by developer35.0%55.1%34.4%45.2%12.5%43.2%64.7%9.2%

More about JobBench

55.4%Reasoning effort: MaxSelf-reported by developer1.4%

More about Humanity's Last Exam

65.7%Reasoning effort: MaxSelf-reported by developer0.0%81.0%

More about SWE-Bench Pro

Can it turn a doctor–patient conversation into an accurate medical record?MedScribeTop 50%#38
83.6%

More about MedScribe

Can it write airtight math proofs?ProofBenchTop 23%#16
75.0%

More about ProofBench

Specialized fields

78.4%Reasoning effort: MaxSelf-reported by developer98.1%85.5%83.4%88.2%81.4%

More about CyberGym

OddBench

57.0%Reasoning effort: High83.5%

By effortNone 54.0%High 57.0%

More about BullshitBench v2

See each evaluation’s detail page for its source and last updateData corrections & removal requests