Tencent Releases Hy4-flash — a 220B-Parameter MoE Model with Open Weights
On November 17, 2026, the tencent/Hy4-flash model for text-generation was published on HuggingFace. It is a MoE architecture with 220.3 billion parameters, of which only 12 billion are active per token. The model accepts text and images, generates text, and supports a context of up to 1 million tokens.
What it means
The full weights in safetensors take up 410.3 GB, with the largest file at 29.5 GB. Running it locally via transformers will require a corresponding amount of VRAM or disk space for offload. The platform already hosts 18 quantized builds in GGUF, MLX, NVFP4, FP8, EXL, and AWQ formats. The most compact working build is Q5 by Baekpica, weighing 98.9 GB.
The Tencent Hy License 1.0 permits use, modification, and distribution of the model, including commercial use, provided authorship notices and a copy of the license are retained. A separate written license is required only for a MaaS business with revenue over 50 million dollars over 12 months or a product with more than 100 million monthly active users. Internal use without third-party access does not require an additional license.
Even in quantized form, the model remains heavy for home hardware: 98.9 GB in Q5 is the level of server GPUs or systems with a large amount of system RAM and offload. For those looking for compact local models, Hy4-flash is unlikely to become an everyday tool, but as an open MoE model with a long context it sets a benchmark for builds on server hardware. You can follow the arrival of lighter quantized variants in the CPU3D news feed.How the method works. The diagram is drawn based on this news note.