Tencent has published the weights of Hy4 flash, its low-cost multimodal model, on Hugging Face, ModelScope and GitCode. The release includes the instruct model, an FP8 version and the pre-trained base model, and the model card now confirms what the launch two days earlier left open: about 220 billion total parameters, roughly 12 billion active per token, a one-million-token context and a custom Tencent licence.
Quick Take
- Released: Hy4 flash, Hy4 flash-FP8 and Hy4 flash-base as open-weight models on Hugging Face, ModelScope and GitCode.
- Size: 220B total parameters including 63.5B “Engram” n-gram embedding parameters; 11.3B active text parameters per token plus 0.43B for vision.
- Context: 1,048,576 tokens; accepts text and images, outputs text.
- Cost: on Tencent’s internal WorkIndex benchmark, average cost per task is 1/13 of Hy4 preview and 1/20 of Hy4, with a score of 42.1 versus 39.7 for Hy4 preview.
- Licence: the Tencent Hy License, which is not a standard open-source licence.
In plain English
Two days ago Tencent launched Hy4 flash as a cheap model you could only use through its cloud. Now anyone can download the model files and run it on their own servers. It is a large model, but only a small part of it works on each word, which keeps it fast and cheap. You still need serious hardware, and the licence has its own rules.
What happened
Tencent’s Hy team added the three repositories under tencent/Hy4-flash with a detailed model card. The independent site ModelFit reported that the weights appeared on Hugging Face on 17 November and are not gated, with the BF16 files totalling about 441 GB and the FP8 files about 287 GB. Tencent provides ready-made vLLM and SGLang container images, a built-in speculative decoding drafter called DFly2 and a finetuning guide.
When i10x covered the TokenHub API launch on 16 November, Tencent’s documents gave no parameter count, licence or date for an open-weight release. The model card now fills those gaps.
How it works
- Mixture of experts: a 42-layer backbone with one dense layer and 41 MoE layers; each token activates 16 of 512 routed experts plus one shared expert.
- Hybrid attention: sparse attention and sliding-window attention are interleaved at a 1:3 ratio, which Tencent says cuts the KV cache to 1/24 of Hy3’s and 1/6 of Hy4’s.
- Engram: learned bigram and trigram embeddings retrieved through hash lookups augment the backbone.
- Speed: megakernels, 4-bit quantization and DFly2 push decoding past 150 tokens per second, according to Tencent.
- User test: in a blind evaluation with 85 internal users across 259 tasks, Hy4 flash scored 2.79 out of 4 for satisfaction against 2.81 for Claude Haiku 5.5, at RMB 0.26 per task.
Why it matters
For developers and companies, Hy4 flash adds a cost-focused multimodal option to the growing pool of large Chinese open-weight models that can be self-hosted for agents and office work.
For inference providers, the licence matters. ModelFit notes that the Tencent Hy License adds commercial terms and requires a separate written licence for model-as-a-service businesses.
For hobbyists, the model is too big for most single machines; an official 4-bit GGUF is about 157 GiB, and llama.cpp support was still an open pull request.
What’s still open
- Independent benchmark results on the public agent tests Tencent cites.
- Transformers support, which Tencent says will be merged into the main branch soon.
- How the licence terms affect hosting by third-party API providers.
Sources
- Tencent Hy team, Hy4 flash model card, Hugging Face, accessed 18 November 2026
- ModelFit, “Tencent Hy4 Flash Weights Are Out: What Hardware Runs It?,” updated 18 November 2026
i10x Perspective
Tencent followed a familiar pattern: launch on its own cloud, then open the weights days later to win developers. With Moonshot’s Kimi K3.1 Pro weights and others, cheap Chinese open models keep raising the floor for what self-hosted AI can do. See our earlier report on the Hy4 flash TokenHub launch and more in the Models section.



