Tencent Releases Hy4 flash: A Cost-Effective Multimodal MoE Model

At a Glance
| Item | Value |
|---|---|
| Repository | tencent/Hy4-flash-base |
| Publisher guide | Tencent: models and licenses |
| Published | 2026-11-17 |
| License | tencent-hy-license |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Tencent has released the “Hy4 flash" series, a multimodal Mixture-of-Experts (MoE) model. It comes in three variants: the instruct version tencent/Hy4-flash, the FP8 quantized version tencent/Hy4-flash-FP8, and the pre-trained-only version tencent/Hy4-flash-base. According to the model card, it is the most cost-efficient model in the Hy series, taking text and image inputs and outputting text.
It is recommended for everyday business tasks (deep search, document processing, slide creation, simple frontend demos, computer use). For complex tasks such as long-term software engineering, the higher-tier Hy4 is recommended.
Specifications
- Architecture: Mixture-of-Experts (MoE). 220B total parameters (including 64B Engram parameters), with 12B active parameters per token.
- Table Breakdown: Total parameters (excluding Engram, ViT, and DFly2) 155.5B, Engram 63.5B, ViT 0.43B, DFly2 0.85B, with per-token active parameters of 11.3B (text) + 0.43B (image).
- Context Length: 1M (1,048,576 tokens).
- Backbone: 42 layers (1 dense layer + 41 MoE layers). Each MoE layer has 512 routed experts and 1 shared expert, activating 16 routed experts per token.
- Attention: Sparse attention and sliding window attention (SWA) arranged in an alternating 1:3 ratio.
- Supports speculative decoding via the built-in 5-layer drafter “DFly2" (0.85B).
- Vision Encoder: Modified SigLIP2-So400m-Patch16-Res512.
Performance
Below are the few-shot evaluations of the pre-trained model (Hy4 flash Base) published by the creators in the model card. These are not third-party verifications. Due to the number of columns, comparisons are limited to Hy4 Base, Hy3 preview Base, DeepSeek-V4-flash-Base, and DeepSeek-V4-Pro-Base, and rows for PIQA, Social IQa, KOR-Bench, C³, and multimodal items have been omitted. Items marked with † are noted in the card as macro-average accuracies for the corresponding text benchmarks.
→ Scroll horizontally to see all columns
| Category | Benchmark | #Shots | Hy4 flash Base | Hy4 Base | Hy3 preview Base | DeepSeek-V4-flash-Base | DeepSeek-V4-Pro-Base |
|---|---|---|---|---|---|---|---|
| Knowledge | MMLU† | 5 | 88.7 | 91.6 | 87.5 | 89.7 | 91.1 |
| MMLU-Pro† | 5 | 71.1 | 73.9 | 65.8 | 68.6 | 73.4 | |
| MMLU-Redux† | 5 | 87.6 | 90.6 | 87.0 | 88.8 | 90.2 | |
| SuperGPQA† | 5 | 57.8 | 62.7 | 52.3 | 52.5 | 54.8 | |
| C-Eval† | 5 | 92.0 | 95.0 | 90.1 | 91.9 | 93.4 | |
| CMMLU† | 5 | 91.6 | 93.1 | 89.6 | 89.8 | 90.7 | |
| SimpleQA | 5 | 37.5 | 57.2 | 26.4 | 31.3 | 52.1 | |
| Chinese-SimpleQA† | 5 | 74.6 | 83.9 | 69.6 | 71.9 | 76.6 | |
| AGIEval (English)† | 3 | 75.4 | 79.3 | 76.0 | 76.3 | 78.2 | |
| AGIEval (Chinese)† | 3 | 90.6 | 91.2 | 87.8 | 88.5 | 89.8 | |
| Reasoning | HellaSwag (3,000-item subset) | 10 | 85.2 | 87.1 | 85.7 | 85.1 | 86.8 |
| ARC-Challenge | 0 | 95.0 | 97.3 | 96.0 | 97.3 | 97.0 | |
| Math | U-MATH | 5 | 69.8 | 71.6 | 60.9 | 57.4 | 64.8 |
| OCW Courses | 5 | 64.0 | 66.7 | 62.6 | 63.3 | 68.2 | |
| Code | MBPP-plus | 3 | 73.7 | 74.0 | 77.4 | 76.8 | 79.1 |
| BigCodeBench (Full) | 0 | 58.8 | 47.9 | 59.6 | 58.0 | 57.5 | |
| Multilingual | Belebele (5-language subset)† | 5 | 93.8 | 95.0 | 94.2 | 94.4 | 95.2 |
| INCLUDE† | 5 | 82.8 | 88.1 | 84.3 | 84.0 | 84.3 |
In this table, Hy4 flash Base outperforms DeepSeek-V4-flash-Base (68.6, 52.5) on MMLU-Pro (71.1) and SuperGPQA (57.8). Its U-MATH score of 69.8 is also second only to Hy4 Base’s 71.6. On the other hand, its SimpleQA score is 37.5, falling well short of Hy4 Base (57.2) and DeepSeek-V4-Pro-Base (52.1). MBPP-plus is 73.7, lower than DeepSeek-V4-flash-Base (76.8) and Hy3 preview Base (77.4). BigCodeBench (Full) at 58.8 is a close match with Hy3 preview Base (59.6) and higher than Hy4 Base (47.9). The higher-tier Hy4 Base outperforms it in most knowledge categories. Additionally, the card notes that exposure to synthetic instruction data during intermediate training may cause few-shot evaluations to skew scores by failing to follow response formats.
For real-world tasks, it scores 42.1 on the creators’ internal benchmark WorkIndex (Hy4 preview is 39.7, Hy4 is 51.4). The average cost per task is 1/13th of Hy4 preview and 1/20th of Hy4. Blind evaluation satisfaction by 85 internal users was 2.79/4, close to Claude Haiku 5.5’s 2.81/4. These are measurements by the creators themselves.
Strengths and Use Cases
The model card recommends everyday business tasks. Specifically, deep search, document processing, slide creation, simple frontend demos, and computer use are listed. The internal benchmark WorkIndex is described as consisting of eight evaluation sets including slide creation/editing, collaborative Markdown writing, document editing, spreadsheet repair, audio/video processing, and visual understanding. It is designed with a focus on quality, output length, and latency for tasks close to actual work.
Meanwhile, for complex tasks like long-term software engineering, the card recommends the higher-tier Hy4. Rather than positioning itself as “the strongest in a specific field," it is oriented towards practicality, prioritizing cost and speed.
It also supports image inputs (tags are multimodal, image-text-to-text). Examples in the card include image description generation. Since the context length is 1M tokens, it is also suited for handling long documents.
Regarding efficiency, the card states the following (all claims are by the creators and have not been verified by third parties):
- The adoption of hybrid attention, LatentMoE, and Engram is reported to reduce the KV cache size to 1/24th of Hy3 and 1/6th of Hy4.
- Reinforcement learning is reported to reduce the average end-to-end output token count by 50%.
- The decoding speed combining megakernel, 4-bit quantization, and DFly2 is reported to exceed 150 tokens/s.
Inference modes are switched via reasoning_effort. Options include "high" (default, thinking enabled) and "no_think" (direct answer). Recommended sampling parameters are temperature=0.9 and top_p=1.0.
Hy4-flash-base is a pre-trained-only base model and is not for chat. Conversation and reasoning examples apply to Hy4-flash and Hy4-flash-FP8. Fine-tuning pipelines and guides are also provided in each repository.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 220.3B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 492GB of VRAM (multi-GPU or CPU offload required) | BF16 | 410.3GB | 492.3GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-18): llama.cpp: not registered, vLLM: not 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, 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 |
|---|---|---|---|---|
| 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) |
| bartowski/Intern-S2-397B-GGUF | 403.4B | — | apache-2.0 | Intern-S2-397B-GGUF Vision-Language Model: ~102GB Memory (2026-09-15) |
How to Get It
Distributed in safetensors format. It is available on Hugging Face, ModelScope, and GitCode. No gating is set, and it can be downloaded on Hugging Face without agreeing to a license. The license is Tencent Hy License (noted as license:other on the card). Details of the terms of use could not be confirmed from the current materials. If you plan commercial use or similar, please check the license text in the repository.
| Model | Description |
|---|---|
tencent/Hy4-flash |
Instruct version |
tencent/Hy4-flash-FP8 |
FP8 quantized instruct version |
tencent/Hy4-flash-base |
Pre-trained base version |
Here is an example download command:
huggingface-cli download tencent/Hy4-flash-FP8
A GGUF version is provided by AngelSlim as AngelSlim/Hy4-flash-GGUF. The card positions this for “end-to-end testing" and not for production operations. Compatibility status with llama.cpp, Ollama, and LM Studio is unverified as it is not mentioned in the materials.
Transformers
Support for Hy4 flash is reported to be “merged soon" into the main branch of transformers. Until then, it must be installed from the following commit:
pip install "transformers[torch,vision] @ git+https://github.com/huggingface/transformers.git@refs/pull/50903/head"
vLLM
For production serving, vLLM or SGLang is recommended. The official prebuilt image for vLLM is vllm/vllm-openai:hy4-flash. The command from the card is as follows:
docker run --gpus all \
-p 8000:8000 \
--ipc=host \
-v ~/.cache/huggingface:/root/.cache/huggingface \
vllm/vllm-openai:hy4-flash tencent/Hy4-flash-FP8 \
--tensor-parallel-size 8 \
--speculative-config '{"num_speculative_tokens":5,"method":"dfly"}' \
--attention-backend FLASHMLA_SPARSE \
--tool-call-parser hy_v4 \
--reasoning-parser hy_v4 \
--enable-auto-tool-choice \
--port 8000 \
--served-model-name hy4-flash
SGLang
The official SGLang image is lmsysorg/sglang:hy4-flash (multi-architecture for x86 and Arm).
docker pull lmsysorg/sglang:hy4-flash
docker run --gpus all \
-p 8000:8000 \
--ipc=host \
lmsysorg/sglang:hy4-flash \
sglang serve \
--trust-remote-code \
--model-path tencent/Hy4-flash-FP8 \
--tp 8 \
--cuda-graph-backend-prefill disabled \
--speculative-algorithm DFLY \
--speculative-dfly-block-size 6 \
--host 127.0.0.1 \
--port 8000
Both examples specify --tensor-parallel-size 8 (--tp 8 in SGLang), assuming a multi-GPU configuration. Once the server starts, you can access it via an OpenAI-compatible API. Speculative decoding via DFly2 is enabled via startup options.
Related Articles
- Qwen3.8-27B-DSpine-b16 Draft Model for Speculative Decoding: 8GB+ VRAM
- Qwen3-8B-DSpine-b16: 4GB+ VRAM
- Hy-Memory-30B-A3B Text Generation Model: 12GB+ VRAM, GGUF Builds
- Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM
What to Read Next
- Find models by VRAM (This model does not fit even in 80GB; its smallest build needs about 492GB) → VRAM quick reference, including models over 80GB
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → Tencent: models, licenses and articles
- Other models for the same task → Other vision-language models
- How to read MMLU, MMLU-Pro, SuperGPQA → Benchmark glossary

