Tencent has added a cheaper member to its Hy model family and, a day after the commercial launch, made it downloadable. The company announced Hy4 flash on November 16, 2026 on its Hy research site and in a corporate press release, describing it as the most cost-effective Hy model so far. Early on November 17 (UTC), the Tencent Hy team published three checkpoints on Hugging Face, ModelScope and GitCode, along with a new GitHub repository, Tencent-Hunyuan/Hy4-flash: the instruction-tuned Hy4 flash, an FP8 version for production serving, and the pre-trained Hy4 flash-base.
Hy4 flash is a mixture-of-experts model that reads text and images, writes text and supports a context window of 1,048,576 tokens. Tencent's launch material gives 224 billion total parameters, including 64 billion so-called Engram parameters, while the model card on GitHub adds up to about 220 billion: a 155.5B backbone, 63.5B of Engram tables, a 0.43B vision encoder and a 0.85B draft model. Either way, only about 12 billion parameters are active for each token.
The design leans heavily on efficiency. Sparse attention and sliding-window attention layers alternate at a ratio of one to three; routed experts work in a compressed latent space, with 16 of 512 experts chosen per token plus one shared expert; and learned bigram and trigram embeddings, retrieved through hash lookups, sit alongside the backbone. A built-in five-layer drafter called DFly2 reuses the main model's cache for speculative decoding. Tencent says these choices shrink the KV cache to one twenty-fourth of Hy3's and one sixth of Hy4's, that reinforcement learning halved average output length, and that megakernels, 4-bit quantization and DFly push decoding past 150 tokens per second.
On pricing, Tencent lists about $0.070 per million input tokens, $0.281 per million output tokens and $0.003 for cached input, far below Hy4. The model is available through Tencent Cloud TokenHub and inside products such as WorkBuddy, Yuanbao, WeCom, QQ and ima, with OpenRouter listed as coming soon. Until November 30 it is free in WorkBuddy and half price on TokenHub.
Tencent's quality claims rest mostly on its own measurements. On WorkIndex, an internal benchmark built from typical WorkBuddy jobs such as building slides, editing documents and repairing spreadsheets, the company reports a score of 42.1 for Hy4 flash against 39.7 for the much larger Hy4 preview, at one thirteenth of the cost per task. In a blind test with 85 internal WorkBuddy users, it says satisfaction came close to Anthropic's Claude Haiku 5.5. The published base-model tables put Hy4 flash-base slightly below DeepSeek-V4-flash-Base on several knowledge tests, and Tencent itself warns that the base model sometimes breaks answer formats in few-shot evaluations. Tencent recommends Hy4 flash for search, document work, slides, simple front-end demos and computer use, and the full Hy4 for long-horizon software engineering.
The weights are not released under a standard open-source licence. The Tencent Hy License builds on Apache 2.0 but adds a Section 10: companies offering the model as a hosted service need a separate written licence once their worldwide revenue exceeds US$50 million in any 12-month period, and any product with more than 100 million monthly users needs one too. Internal use is exempt. Writing for the model router OrcaRouter on November 18, jinhao.song argued that this clause, and not the benchmark tables, is how the release will be judged, and noted that no independent benchmark score was yet available.
Serving support arrived at launch but is still pinned. The vLLM project's recipe relies on a dedicated container image rather than a released package, targets NVIDIA Hopper and Blackwell data-centre GPUs, offloads the Engram tables to host memory and does not yet cover AMD hardware. For teams that want a long-context multimodal model they can run themselves, Hy4 flash is a serious new option; whether its efficiency claims hold up will depend on outside testing.