Hy4 flash: Tencent Hunyuan's low-cost multimodal MoE model

Instruction model uploaded by the Tencent Hunyuan team on 11 17. The model card states a total parameter count of 220B (of which approximately 64B are N-gram embedding Engrams), with approximately 11.3B text parameters and 0.43B visual parameters activated per token. It natively accepts text and image inputs and outputs text, with a context window of 1M tokens. Compared to Hy4, it alternates between sparse attention and sliding window attention at a ratio of 1:3, places routed experts in a low-dimensional latent space (selecting 16 out of 512, plus 1 shared experts), and incorporates binary/ternary N-gram embeddings. It also includes a speculative decoding draft module, DFly2, with 0.85B parameters and 5 layers. An FP8 version and the pre-trained base model Hy4 flash-base are also released, under the Tencent Hunyuan License.

Original title: tencent/Hy4-flash

Instruction model uploaded by the Tencent Hunyuan team on 11 17. The model card states a total parameter count of 220B (of which approximately 64B are N-gram embedding Engrams), with approximately 11.3B text parameters and 0.43B visual parameters activated per token. It natively accepts text and image inputs and outputs text, with a context window of 1M tokens. Compared to Hy4, it alternates between sparse attention and sliding window attention at a ratio of 1:3, places routed experts in a low-dimensional latent space (selecting 16 out of 512, plus 1 shared experts), and incorporates binary/ternary N-gram embeddings. It also includes a speculative decoding draft module, DFly2, with 0.85B parameters and 5 layers. An FP8 version and the pre-trained base model Hy4 flash-base are also released, under the Tencent Hunyuan License.

Why it matters Officially, it scores slightly higher than Hy4 preview on office tasks, with a per-task cost only one-thirteenth as high; it includes a built-in draft module to accelerate long-context decoding

Upvotes 213 · Likes 234 · Downloads 159

Reasons

  • Officially, the KV cache is reduced to 1/24 of Hy3 and 1/6 of Hy4, and reinforcement learning reduces average output tokens by 50%; combined with megakernels, 4-bit quantization, and DFly2, decoding speed exceeds 150 tokens/s
  • On the official internal benchmark WorkIndex (covering 8 task groups such as creating slides, editing documents, fixing tables, and audio/video processing), it scores 42.1, compared to 39.7 for Hy4 preview and 51.4 for Hy4; the average cost per task is approximately 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017
  • Blind evaluation by 85 internal WorkBuddy users across 259 tasks: satisfaction score 2.79/4 (Claude Haiku 5.5 scored 2.81/4), median completion time 6.7 minutes, and approximately 0.26 yuan per task

Usage

  • Transformers support has not yet been merged; install from transformers PR #50903 using the commands in the model card. Load using AutoModelForImageTextToText, setting reasoning_effort to high (default, enables thinking) or no_think. Recommended temperature is 0.9 and top_p is 1.0
  • Service deployment: Use the official vLLM image vllm/vllm-openai:hy4-flash to run the FP8 version (example: --tensor-parallel-size 8, DFly speculative decoding for 5 tokens, --tool-call-parser hy_v4 --reasoning-parser hy_v4); for SGLang, use lmsysorg/sglang:hy4-flash (--tp 8, --speculative-algorithm DFLY)
  • Officially recommended for deep search, document processing, slide creation, simple frontend demos, and computer operation; for long-chain software engineering, Hy4 is recommended instead. AngelSlim also provides a GGUF version for end-to-end testing, and the repository includes a fine-tuning guide

Risks

  • The Tencent Hunyuan License is not a standard open-source license: if you operate a Model-as-a-Service (MaaS) business and your global annual revenue exceeds 5000 USD or your product's monthly active users exceed 1 hundred million, you must sign a separate written license (except for purely internal use)
  • All official deployment examples use 8-card tensor parallelism; memory requirements for single-card or consumer-grade GPUs are not specified
  • WorkIndex and blind evaluation results are official internal data; the official team notes that the base model occasionally fails to follow formatting in few-shot evaluations, which may affect scores. This entry has not been independently verified

Open

Hugging Face Trending

Explore the latest AI Radar updates