Tencent Hy4 is Tencent's open-weight AI model, released 03/11/2026. Tencent's Hy4 model card publishes few-shot results for the Hy4 base model as text and shows the instruct-model benchmark charts only as images, which could not be read. The text-readable official figures are the base-model table (Hy4 Base leads Hy4 flash Base, Hy3 preview Base and both DeepSeek V4 base models on most rows), a 579-task finance evaluation where Hy4 scores 3.04 out of 4, and WorkIndex 51.4. The WorkIndex sub-tests come from our Hy4 flash article citing Tencent's flash card. Instruct-model coding and agent benchmark rows are not included because their values could not be read. The release date is the Hugging Face repository creation date.
Visit Tencent Hy4AI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does Tencent Hy4 score on independent benchmarks?
Tencent Hy4 ranks #17 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 34.6, as measured on 22 November 2026.
Where does Tencent Hy4 rank against other models?
How does Tencent Hy4 do on individual tests?
| Evaluation (run by Artificial Analysis) | Tencent Hy4 | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 37.4% | 67.2% |
| FrontierCode 2.0 | 31.4% | 48.8% |
| Humanity's Last Exam | n/a | 74.3% |
| GDPval-AA (Elo) | n/a | 1866 |
How has Tencent Hy4's score changed over time?
We have tracked Tencent Hy4 on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the Tencent Hy4 system card report?
Tencent's Hy4 model card publishes few-shot results for the Hy4 base model as text and shows the instruct-model benchmark charts only as images, which could not be read. The text-readable official figures are the base-model table (Hy4 Base leads Hy4 flash Base, Hy3 preview Base and both DeepSeek V4 base models on most rows), a 579-task finance evaluation where Hy4 scores 3.04 out of 4, and WorkIndex 51.4. The WorkIndex sub-tests come from our Hy4 flash article citing Tencent's flash card. Instruct-model coding and agent benchmark rows are not included because their values could not be read. The release date is the Hugging Face repository creation date.
These are Tencent's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Sources: Hy4 (Hugging Face model card) (29/10/2026), Hy4 flash (Hugging Face model card) (16/11/2026). Last checked 22/11/2026.
Knowledge & science
| Benchmark | Tencent Hy4 | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base |
|---|---|---|---|---|
| MMLU (Base)5-shot; macro-averaged accuracy; base model | 91.6% | 88.7% | 87.5% | 89.7% |
| MMLU-Pro (Base)5-shot; macro-averaged accuracy; base model | 73.9% | 71.1% | 65.8% | 68.6% |
| MMLU-Redux (Base)5-shot; macro-averaged accuracy; base model | 90.6% | 87.6% | 87% | 88.8% |
| SuperGPQA (Base)5-shot; macro-averaged accuracy; base model | 62.7% | 57.8% | 52.3% | 52.5% |
| C-Eval (Base)5-shot; macro-averaged accuracy; base model | 95% | 92% | 90.1% | 91.9% |
| CMMLU (Base)5-shot; macro-averaged accuracy; base model | 93.1% | 91.6% | 89.6% | 89.8% |
| SimpleQA (Base)5-shot; base model | 57.2% | 37.5% | 26.4% | 31.3% |
| Chinese-SimpleQA (Base)5-shot; macro-averaged accuracy; base model | 83.9% | 74.6% | 69.6% | 71.9% |
| AGIEval English (Base)3-shot; macro-averaged accuracy; base model | 79.3% | 75.4% | 76% | 76.3% |
| AGIEval Chinese (Base)3-shot; macro-averaged accuracy; base model | 91.2% | 90.6% | 87.8% | 88.5% |
Reasoning & maths
| Benchmark | Tencent Hy4 | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base |
|---|---|---|---|---|
| HellaSwag (Base)10-shot; 3,000-item subset; base model | 87.1% | 85.2% | 85.7% | 85.1% |
| PIQA (Base)4-shot; base model | 85.3% | 84.4% | 84.6% | 83.8% |
| ARC-Challenge (Base)0-shot; 299-item subset; base model | 97.3% | 95% | 96% | 97.3% |
| Social IQa (Base)0-shot; base model | 82.7% | 80.1% | 80.5% | 72.5% |
| KOR-Bench (Base)3-shot; base model | 57.6% | 54.7% | 54.6% | 53.4% |
| C3 (Base)0-shot; base model | 82.4% | 77.3% | 75.9% | 74.1% |
| U-MATH (Base)5-shot; base model | 71.6% | 69.8% | 60.9% | 57.4% |
| OCW Courses (Base)5-shot; base model | 66.7% | 64% | 62.6% | 63.3% |
Coding
| Benchmark | Tencent Hy4 | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base |
|---|---|---|---|---|
| MBPP-plus (Base)3-shot; base model | 74% | 73.7% | 77.4% | 76.8% |
Other
| Benchmark | Tencent Hy4 | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base |
|---|---|---|---|---|
| Belebele, 5-language subset (Base)5-shot; macro-averaged accuracy; 4,475 examples; base model | 95% | 93.8% | 94.2% | 94.4% |
| INCLUDE, 5,000-example Lite subset (Base)5-shot; macro-averaged accuracy; 44 languages; base model | 88.1% | 82.8% | 84.3% | 84% |
| Finance workflow evaluation (instruct)579 real-world finance tasks, average score out of 4; Hy4 instruct model | 3.04 /4 | n/a | n/a | n/a |
| WorkIndex (Tencent internal)Tencent internal set of eight office-task tests, run in WorkBuddy; figure from the Hy4 flash model card | 51.4 score | n/a | n/a | n/a |
Multimodal
| Benchmark | Tencent Hy4 | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base |
|---|---|---|---|---|
| MMMU-Pro Standard, 10 options (Base)4-shot; base model | 59.6% | 52% | n/a | n/a |
| SimpleVQA (Base)4-shot; base model | 54.5% | 48.5% | n/a | n/a |
| CVBench (Base)4-shot; base model | 86.7% | 83.4% | n/a | n/a |
| RealWorldQA (Base)4-shot; base model | 72.3% | 71.1% | n/a | n/a |
| TextVQA (Base)4-shot; base model | 78.2% | 73.4% | n/a | n/a |
| CountBenchQA (Base)4-shot; base model | 90.6% | 86.6% | n/a | n/a |
Agents & tool use
| Benchmark | Tencent Hy4 | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base |
|---|---|---|---|---|
| PPTGenBench (slide creation)WorkIndex test; run in the WorkBuddy command-line tool | 61.5 score | n/a | n/a | n/a |
| PPTCoEditBench (slide editing)WorkIndex test; run in the WorkBuddy command-line tool | 78.9 score | n/a | n/a | n/a |
| PPTFixBench (slide repair)WorkIndex test; run in the WorkBuddy command-line tool | 45.6 score | n/a | n/a | n/a |
| MdCoEditBench (Markdown co-editing)WorkIndex test; run in the WorkBuddy command-line tool | 57.7 score | n/a | n/a | n/a |
| WordEditBench (document editing)WorkIndex test; run in the WorkBuddy command-line tool | 57.2 score | n/a | n/a | n/a |
| ExcelFixBench (spreadsheet repair)WorkIndex test; run in the WorkBuddy command-line tool | 27.8 score | n/a | n/a | n/a |
| AudioVideoBenchWorkIndex test; run in the WorkBuddy command-line tool | 27.4 score | n/a | n/a | n/a |
| BuddyVQA Atlas (image questions)WorkIndex test; run in the WorkBuddy command-line tool | 55.5 score | n/a | n/a | n/a |
What did safety testing find for Tencent Hy4?
Safety level: No vendor safety rating on the model card
- The Hy4 model card text lists benchmarks, licence and inference settings; no safety evaluation was found. The instruct benchmark images, including a limitations chart, could not be read.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page.
Tencent Model Timeline
1,048,576 tokens (1M) context
131k tokens context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
