Hy4
Licence: Tencent Hy License — Apache-derived; section 10 requires a separate written licence from model-as-a-service operators above US$50M revenue or 100M monthly users. Internal use exempt.
Tencent's flagship MoE, weights published Nov 20, 2026 on Hugging Face, ModelScope and GitCode (instruct ~1.55TB, FP8 ~0.79TB, base ~1.54TB, plus a system card): 770B total parameters in the language backbone with 49B active per token, 78 layers (1 dense, 77 MoE) with 256 routed experts of which 8 fire plus 1 shared expert, a 1,048,576-token context, native vision (1.23B encoder) and computer use, and a built-in 3.3B DFly2 speculative-decoding drafter that shares the model's KV cache. Memory by arithmetic: FP8 weights alone are about 770 GB before KV cache — an 8-GPU node of H200/B200 class cards is the serving unit (Tencent's own vLLM and SGLang recipes use tensor parallelism across 8 GPUs); at BF16 the weights are about 1.54 TB, a multi-node job. Watch-outs: the licence's section 10 gate applies to anyone reselling Hy4 inference as a service at scale (internal use is exempt); all benchmarks are Tencent's own evaluations — no independent evaluation exists yet. Hy4 flash (220B/12B) is the smaller sibling, released Nov 17.
What it needs
How we work this out: memory = the model's total parameters × bytes per weight — 2 bytes at full precision (FP16), 1 byte at 8-bit, half a byte at 4-bit. Weights only — add room for the conversation memory (KV cache) and the serving software. A GPU counts as running a build when its memory covers the weights plus about 20% headroom. Real figures move with context length and software.
GPUs that can run it
Ranked by how well the model runs, then by price.
No single GPU can run Hy4 — even its 4-bit build needs 385GB, plus room to work, and no single card holds that. Needs at least 2× B300 by memory at 4-bit — a memory fact, not a performance promise. The prices board shows the biggest cards and what each is short by, with the single-GPU and node prices a setup like this is built from.
The minimum setup, by memory
Needs at least 2× B300 by memory at 4-bit — a memory fact, not a performance promise.
No single card runs Hy4 alone at this size — it is a multi-GPU job. The board lists the cards, nodes and racks a setup is built from, with node offers shown by their total price.
Checked Nov 21, 2026
Select a GPU See these GPUs on the prices board, ranked for this model.