Tencent logo

Hy4 flash

Multimodal
Tencent

Hy4 flash is Tencent Hy's most cost-effective Hy model, announced on 2026-11-16 (API id hy4-flash). It has 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters per token. Like Hy4, it is natively multimodal and has a 1,048,576-token context window (Tencent Cloud TokenHub documents 960K max input and 64K max output). Official TokenHub USD list pricing, from Tencent Cloud documentation https://www.tencentcloud.com/document/product/1300/78937, is $0.070 per million input tokens, $0.281 per million output tokens, and $0.003 per million cached input tokens. That same pricing notice offers 50% off Hy4 flash list API pricing until 12:00 Beijing time on 2026-11-30 (11:00 PM ET on 2026-11-29). The official research page also says WorkBuddy access is free until 30 November. Tencent says Hy4 flash will launch on OpenRouter soon and that it plans to open-source the model; no release date was given for the weights, and neither the OpenRouter listing nor the weights were published at catalog time. Tencent's evaluation emphasizes its internal WorkIndex benchmark and a WorkBuddy blind test, and the launch page also reports self-reported scores on public agent and multimodal benchmarks in an official appendix, evaluated at the highest available reasoning setting.

Key Specifications

Parameters
224.0B
Context
960.0K
Release Date
November 16, 2026
Average Score
51.1%

Timeline

Key dates in the model's history
Announcement
November 16, 2026
Last Update
November 18, 2026
Today
November 20, 2026

Technical Specifications

Parameters
224.0B
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$0.07
Output (per 1M tokens)
$0.28
Max Input Tokens
960.0K
Max Output Tokens
64.0K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
APEX-Agents
APEX-Agents v1.1 pass@1 score 63.1 for Hy4 flash, from the official research-page appendix. Official notes: official reference agent harness (Loop), maximum of 250 steps per trial. Highest available reasoning setting. • Self-reported
63.1%
AutomationBench v1.0.6
AutomationBench v1.0.6 score 52.2 for Hy4 flash, from the official research-page appendix. Official notes: 600 public tasks released by Zapier, evaluated according to the official method. Highest available reasoning setting. • Self-reported
52.2%
DeepSWE 1.1
DeepSWE v1.1 score 64.6 for Hy4 flash, from the official research-page appendix. Official notes: mini-SWE-agent; each task runs in a sandbox with 8 CPU cores and 16 GB of memory. Highest available reasoning setting. • Self-reported
64.6%
Humanity's Last Exam (with tools)
HLE-Full with tools score 55.2 for Hy4 flash, from the official research-page appendix (row labeled HLE-Full (with tools)). Highest available reasoning setting. • Self-reported
55.2%
MMMU-Pro
MMMU-Pro score 75.8 for Hy4 flash, from the official research-page appendix. The appendix gives no further MMMU-Pro footnote. Highest available reasoning setting. • Self-reported
75.8%
OSWorld 2.1 (partial)
OSWorld 2.1 full set, partial score 68.1 for Hy4 flash, from the official research-page appendix (row labeled OSWorld 2.1 (Full, partial score)). Filed under osworld-2.1-partial, the catalog id for OSWorld 2.1 partial-credit scores. Highest available reasoning setting. • Self-reported
68.1%
Program Bench
ProgramBench Almost Solved score 16.0 for Hy4 flash, from the official research-page appendix. Official notes: Almost Solved score using mini-SWE-agent. Highest available reasoning setting. • Self-reported
16.0%
Terminal-Bench 4.0
Terminal-Bench 4.0 score 19.2 for Hy4 flash, from the official research-page appendix. Official notes: Claude Code harness, up to 500 turns and a 13-hour timeout per trial; resources capped at 16 CPUs and 32 GB RAM, with the three GPU-dependent tasks additionally receiving a single H100 GPU. Highest available reasoning setting. • Self-reported
19.2%
ZEROBench
ZeroBench-main score 45.8 for Hy4 flash, from the official research-page appendix (row labeled ZeroBench-main (with tools, pass@5)). Official notes: network-isolated Jupyter environment with image outputs as the tool interface, maximum of 80 tool-calling rounds per run; each problem is evaluated over 10 independent runs to estimate pass@5. Highest available reasoning setting. • Self-reported
45.8%

License & Metadata

License
unknown
Announcement Date
November 16, 2026
Last Updated
November 18, 2026

Similar Models

All Models

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.