At launch on November 16, Hy4 Flash shipped without a model card, a license or a weights repository — OrcaRouter's launch check listed the gaps plainly. Two days later, on November 18, Tencent closed most of that list in one move: the Hy4 Flash weights are open, in three variants, with the model card and license published alongside. The speed is impressive; the contents deserve a closer read. The license is not Apache 2.0, the model card's parameter total does not match the launch release, and the base model's pretraining scores sit below Hy4 almost across the board.
The release covers three lines: Hy4 flash, the instruction-tuned release; Hy4 flash-FP8, the 8-bit quantised version; and Hy4 flash-base, the pretrained base model. All four channels carry them at once: Hugging Face (tencent/Hy4-flash), ModelScope, GitCode and GitHub. We checked the weight sizes against the Hugging Face API: the standard weights total 440.5 GB across 148 safetensors shards; the FP8 version is 287.1 GB and the base version 438.8 GB. Each variant has its own job: the tuned release is ready to use, the FP8 build saves memory, and the base model is there for anyone who wants to do their own post-training.
The license is the easiest place to trip up. Hy4 Flash does not use the Apache 2.0 of the Hy4 preview; it ships under Tencent's own Tencent Hy License 1.0, which is derived from Apache 2.0 with added commercial terms. The key clause is Article 10: MaaS (model-as-a-service) operators with over $50 million in annual revenue, or products with over 100 million monthly active users, need a separate written license from Tencent before use. Ordinary use by individuals and small teams is not caught by that clause — but large platforms must read the license text and go through the written process before commercial use. Reading "Tencent open-sourced it" as "Apache 2.0, use it however you like" will go wrong here.
The second discrepancy worth noting is the parameter count. The November 16 launch release puts Hy4 Flash at 224B total parameters; the November 18 model card lists 220B. The model card's own parameter table adds up to the 220B figure: a 155.5B backbone, 63.5B of Engram, a 0.43B ViT vision encoder and the 0.85B DFly2 speculative-decoding draft model — about 220.3B in total. The 4B gap is unexplained by Tencent: it may be a counting difference, or the launch figure may simply be a rounded number. When citing the size, name the source — 220B for the model card, 224B for the launch release — and do not mix the two.
One detail in the model card worth singling out is DFly2: a 0.85B speculative-decoding draft model shipped inside the weights package, paired with the main model to speed up decoding. On deployment, the card includes vLLM and SGLang container images plus an 8-way tensor-parallel deployment example — enough to get the model running across several GPUs without assembling the setup from scratch.
The model card includes a pretraining benchmark table for Hy4 flash Base, measured few-shot and shown alongside Hy4 Base, Hy3 preview Base and two DeepSeek-V4 base models. The key rows, all Tencent-reported:
| Test (model card, few-shot base evaluation, Tencent-reported) | Hy4 flash Base |
|---|---|
| MMLU | 88.7 |
| MMLU-Pro | 71.1 |
| C-Eval | 92.0 |
| CMMLU | 91.6 |
| SimpleQA | 37.5 |
| MMMU-Pro | 52.0 |
Read across, the Flash base sits below Hy4 Base on almost every item — which fits its positioning: the lightweight version is not a shrunken flagship; its base capability is a tier lower. The model card carries its own disclaimer: the base model saw large amounts of synthetic instruction data during mid-training, and failures to follow the few-shot format may distort the scores. In other words, treat these numbers as a rough read on base capability, not as the tuned release's results; the card gives no instruction-following scores for the tuned Hy4 flash itself.
The same day brought one more Tencent-side claim: Hy4 Flash handled 500 trillion tokens on its first day live. KuCoin, citing Science and Technology Daily, attributes the figure to an official Tencent disclosure on November 18. We cannot independently verify the counting basis (whether cached or deduplicated tokens are included), so keep the "Tencent says" qualifier when citing it.
Flash is open; the flagship is not moving. The full Hy4 release's weights are still unpublished as of November 18, with Tencent's wording still "accelerating" and no date attached. The two models' open-source timelines have now split: the preview open-sourced on launch day, Flash followed two days later, and the full release's promise is still waiting to land.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against Hugging Face — tencent/Hy4-flash model card (Nov 18, 2026)
← Back to the Tencent Hunyuan Hy4 topic page