Tencent has open-sourced its Hy4 Flash model family in three versions, including an FP8 build designed to reduce the memory needed to run it. The release gives developers a production-oriented model, a lower-memory option and a base model they can continue training, though this is no small model to tuck into a spare laptop.
Tencent AI Watch analysis
What happened
Hy4 Flash is the instruction-tuned version for dialogue, reasoning and agent tasks. Tencent has also released Hy4 Flash-FP8, an 8-bit quantised version intended to reduce VRAM use, and Hy4 Flash Base, which is not instruction-tuned and can be adapted through further training.
The model has 224 billion parameters, with 12 billion active during each inference, according to a report by ChainCatcher carried by KuCoin. The report says the models are available through GitHub, Hugging Face, ModelScope and GitCode. It also says the weights use Tencent’s Hy License, with conditions including separate authorisation for API providers above specified revenue or user-scale thresholds.
Why it matters
The three versions serve different jobs: use the instruction-tuned model directly, try FP8 to ease deployment demands, or start from the base weights for further training. That makes the release more useful than a single set of weights, particularly for developers weighing customisation against hardware requirements.
The FP8 label is a relative relief, not a magic spell. A 224-billion-parameter model still carries substantial deployment demands, and the licence conditions matter for some commercial uses.
Our read
There is meaningful choice here for developers, from a ready-to-use model to weights for further work. Check the licence and your hardware budget before building around it; “open-source” does not mean “no conditions” or “runs on anything”.
What to watch
- Whether developers publish usable deployment guidance for the FP8 version. – How the three versions perform in independent evaluations and real workloads. – Whether Tencent publishes the promised Hy4 release, for which the report says no date has been announced.
Discussion spark: Would you choose a large model with ready-to-use, FP8 and base versions, or does the hardware burden and licence make a smaller alternative more practical?
Sources and evidence
- Tencent HuanYuan Open-Sources Hy4 Flash with Three Model Versions – KuCoin (18 November 2026, 06:44 UTC)
- Tencent Hunyuan Opens Source Hy4 Flash Models – KuCoin (18 November 2026, 06:44 UTC)
Independent WittyWires tracker for public updates about Tencent AI. Not affiliated with or endorsed by Tencent AI; this is not an official account.