Hy4 Model Specifications, Benchmarks, and Use Cases

At a Glance
| Item | Value |
|---|---|
| Repository | tencent/Hy4-base |
| Family guide | Hy4 guide (2 articles) |
| Publisher guide | Tencent: models and licenses |
| Published | 2026-11-18 |
| License | tencent-hy-license |
| Formats | safetensors |
| Paper | arXiv:2512.02556, arXiv:2603.12201 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Specifications
- Architecture: Mixture-of-Experts (MoE)
- Total Parameters: 770B (excluding ViT, adapters, and DFly2)
- Active Parameters per Token: Text 49B + Vision 1.23B
- ViT Parameters: 1.23B
- DFly2 Parameters: 3.3B
- Context Length: 1M (1,048,576 tokens)
- Backbone Layers: 78 layers (1 dense FFN layer, 77 MoE FFN layers)
- Hidden Size: 6,144
- Vocabulary Size: 120,832
- Attention Structure: Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache
Performance
The benchmark comparison results for the pretrained models published by the model creators in the model card are as follows:
→ Scroll horizontally to see all columns
| Category | Benchmark | #Shots | Hy4 Base | Hy4 flash Base | Hy3 preview Base | DeepSeek-V4-flash-Base | DeepSeek-V4-Pro-Base |
|---|---|---|---|---|---|---|---|
| Knowledge | MMLU† | 5 | 91.6 | 88.7 | 87.5 | 89.7 | 91.1 |
| MMLU-Pro† | 5 | 73.9 | 71.1 | 65.8 | 68.6 | 73.4 | |
| MMLU-Redux† | 5 | 90.6 | 87.6 | 87.0 | 88.8 | 90.2 | |
| SuperGPQA† | 5 | 62.7 | 57.8 | 52.3 | 52.5 | 54.8 | |
| C-Eval† | 5 | 95.0 | 92.0 | 90.1 | 91.9 | 93.4 | |
| CMMLU† | 5 | 93.1 | 91.6 | 89.6 | 89.8 | 90.7 | |
| SimpleQA | 5 | 57.2 | 37.5 | 26.4 | 31.3 | 52.1 | |
| Chinese-SimpleQA† | 5 | 83.9 | 74.6 | 69.6 | 71.9 | 76.6 | |
| AGIEval (English)† | 3 | 79.3 | 75.4 | 76.0 | 76.3 | 78.2 | |
| AGIEval (Chinese)† | 3 | 91.2 | 90.6 | 87.8 | 88.5 | 89.8 | |
| Reasoning | HellaSwag (3,000-item subset) | 10 | 87.1 | 85.2 | 85.7 | 85.1 | 86.8 |
| PIQA | 4 | 85.3 | 84.4 | 84.6 | 83.8 | 84.9 | |
| ARC-Challenge | 0 | 97.3 | 95.0 | 96.0 | 97.3 | 97.0 | |
| Social IQa | 0 | 82.7 | 80.1 | 80.5 | 72.5 | 73.1 | |
| KOR-Bench | 3 | 57.6 | 54.7 | 54.6 | 53.4 | 55.4 | |
| C³ | 0 | 82.4 | 77.3 | 75.9 | 74.1 | 78.3 | |
| Math | U-MATH | 5 | 71.6 | 69.8 | 60.9 | 57.4 | 64.8 |
| OCW Courses | 5 | 66.7 | 64.0 | 62.6 | 63.3 | 68.2 | |
| Code | MBPP-plus | 3 | 74.0 | 73.7 | 77.4 | 76.8 | 79.1 |
| Multilingual | Belebele (5-language subset)† | 5 | 95.0 | 93.8 | 94.2 | 94.4 | 95.2 |
| INCLUDE† | 5 | 88.1 | 82.8 | 84.3 | 84.0 | 84.3 | |
| Multimodal | MMMU-Pro (Standard, 10 opts) | 4 | 59.6 | 52.0 | — | — | — |
| SimpleVQA | 4 | 54.5 | 48.5 | — | — | — | |
| CVBench | 4 | 86.7 | 83.4 | — | — | — | |
| RealWorldQA | 4 | 72.3 | 71.1 | — | — | — | |
| TextVQA | 4 | 78.2 | 73.4 | — | — | — | |
| CountBenchQA | 4 | 90.6 | 86.6 | — | — | — |
Additionally, score comparisons (out of 4) in practical financial tasks are also presented.
| Model | Average Score (out of 4) |
|---|---|
| Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
According to the evaluation data shown in the model card, Hy4 Base demonstrates extremely high capabilities in knowledge, multilingualism, and factual verification. Specifically, it achieves the highest scores among the listed models in numerous categories, including MMLU (91.6) for general knowledge, MMLU-Pro (73.9) for reasoning, SuperGPQA (62.7) for graduate-level expertise, C-Eval (95.0) for Chinese knowledge, and SimpleQA (57.2) for accurate fact-based answering. On the other hand, its score on MBPP-plus, which evaluates programming tasks, is 74.0, falling below DeepSeek-V4-Pro-Base (79.1) and DeepSeek-V4-flash-Base (76.8). From this comparison, it can be said that while it holds a strong advantage in broad academic knowledge, fact answering, and financial workflow processing, it sometimes falls short of other models in specific coding task evaluations.
Strengths and Use Cases
Hy4 is a multimodal model that integrates the understanding of text, images, and video, supporting a wide range of practical and research workflows. The main strengths and intended uses presented in the model card are as follows:
Vision and Computer Use
By combining visual perception, reasoning, and tool use, it can perform detailed image analysis and interpret complex charts. It supports computer operation across command-line interfaces (CLI) and graphical user interfaces (GUI), making it proficient in tasks using specialized software such as Blender and CAD, as well as planning and executing long workflows spanning multiple applications.
Front-End and Game Development
In front-end development, it is optimized for UI interaction, business logic, state management, and full-stack integration to generate interfaces that are complete both in visual polish and functionality. In game development, it supports feature implementation, gameplay logic, state tracking, and asset pipeline construction. Besides adapting to major game engines such as Unreal Engine, Unity, and Godot, it can also be used for multiplayer game development.
Financial and Office Workflows
It is suitable for a wide variety of financial tasks, such as fund research, equity analysis, quantitative strategies, asset allocation, corporate finance, and tax accounting. Enhanced search capabilities improve timely information retrieval, multi-source integration, and fact-verification accuracy, aiming to reduce hallucinations and reasoning errors. Furthermore, it smoothly handles real-time co-editing and formatting adjustments for office documents such as Word, Excel, and PowerPoint.
Scientific Research and AI R&D
It strongly supports research workflows by integrating deep reasoning, scientific computation, and the execution of long tasks. It can handle tasks across multiple specialized domains, such as mathematical construction, physical modeling, earth science, systems engineering, and socio-economic decision-making. It is also utilized for optimizing in-house pretraining infrastructure and conducting the model’s autonomous learning experiments.
Long-Context Personal Assistant
By providing relevant context, it can comprehensively organize information from past chat histories, emails, documents, calendars, and meeting notes. It supports tasks such as tracking project history and updates, extracting action items, creating meeting briefs, notifying deadlines, and prioritizing tasks.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 771.1B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 1724GB of VRAM (multi-GPU or CPU offload required) | BF16 | 1436.4GB | 1723.7GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-20): llama.cpp: not registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Not usable in Ollama, LM Studio and llama.cpp yet.
The publisher ships safetensors only, and llama.cpp’s registry does not list this architecture. llama.cpp would need to add support before these tools can run it. Today it can be run with transformers or vLLM, using the memory figures in the table above.
License — tencent-hy-license: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (tencent-hy-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Recent Models in the Same Size Class
Models with over 40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| tencent/Hy4-flash-base | 220.3B | — | tencent-hy-license | Tencent Releases Hy4 flash: A Cost-Effective Multimodal MoE Model (2026-11-18) |
| nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | 124.5B | 80GB | openmdw-1.1 | NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16: 80GB+ VRAM, GGUF Builds (2026-11-17) |
| deepseek-ai/DeepSeek-V4.1-Flash-1017 | 763.2B | — | mit | DeepSeek-V4.1-Flash-1017 Vision-Language Model: ~570GB Memory (2026-11-17) |
| Qwen/Qwen3.8-Flash-Next | 180.0B | — | qwen-community-1.0 | Qwen3.8-Flash-Next Multimodal MoE Model: ~402GB Memory (2026-10-04) |
| ggml-org/GLM-5.3-Flash-GGUF | 321.3B | — | other | GLM-5.3-Flash-GGUF Vision-Language Model: ~150GB Memory (2026-10-03) |
How to Get It
Hy4 model weights are available on Hugging Face, ModelScope, and GitCode. The pretrained base model tencent/Hy4-base and the FP8-quantized instruction-following model tencent/Hy4-FP8 are distributed in safetensors format, alongside GGUF compressed versions by AngelSlim. Usage is subject to the proprietary Tencent Hy License.
For deployment environments, inference frameworks such as vLLM and SGLang are recommended. Servers can be launched using official Docker images (vllm/vllm-openai:hy4 or lmsysorg/sglang:hy4), allowing access via OpenAI-compatible APIs.
When making API calls, you can use the reasoning_effort parameter to set the thinking mode. You can choose between a deep thinking mode suitable for complex tasks like math and coding ("high", the default value) and a quick-response mode ("no_think"). Recommended generation parameters are temperature=0.9 and top_p=1.0.
Related Articles
- Tencent Releases Hy4 flash: A Cost-Effective Multimodal MoE Model
- Qwen3.8-27B-DSpine-b16 Draft Model for Speculative Decoding: 8GB+ VRAM
- Tencent Releases WeVisOne-VL Vision Language Models
- Qwen3-8B-DSpine-b16: 4GB+ VRAM
What to Read Next
- Find models by VRAM (This model does not fit even in 80GB; its smallest build needs about 1724GB) → VRAM quick reference, including models over 80GB
- Explore the same model family → Hy4 family overview (2 articles, 3 converted builds)
- Engines that run this model → vLLM
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → Tencent: models, licenses and articles
