Hy3
Tencent finally drops the community license — mind which checkpoint you grab
What it is
Tencent Hunyuan's 295B-total / 21B-active MoE reasoner (192 experts, top-8, plus a 3.8B multi-token-prediction layer) with 256K context — released under plain Apache-2.0, a first for the Hunyuan line. Official 1-bit and 4-bit GGUF builds put a flagship-scale model on a single GPU via llama.cpp.
Why it's interesting
GLM-5.2-class results at less than half the total parameters — 78.0 SWE-Bench Verified, 90.4 GPQA-Diamond — and Tencent's switch from its restrictive community license to Apache-2.0 is a genuine licensing policy shift worth rewarding with attention.
Use cases
- Cost-efficient reasoning and agent backends
- Near-frontier local inference on a single big GPU via official GGUFs
- Fine-tuning without license anxiety
Who it's for
Teams priced out of trillion-parameter hosting; local-AI enthusiasts
Setup
Moderate. FP8/BF16 want a multi-GPU node; the official 1-bit/4-bit GGUFs with MTP run on one GPU under llama.cpp
Limitations & cautions
Coding trails GLM-5.2 in head-to-heads, the release lineage is brand new with little independent evaluation, and the aggressive 1-bit quantizations cost real quality.
Editorial takeaway
Every Hunyuan release used to come with a map of where you couldn't use it. This one just comes with weights.