MODELS →

Hy4 flash

Size
220.3B
12B active — only part of the model works on each answer, but all of it has to fit in memory
Context
1M
the longest conversation the model is built to hold
Lab
Tencent
the lab that released it
Released
Nov 17, 2026

Licence: Tencent Hy License v1.0 — an Apache 2.0 derivative: free use, modification and commercial use, but a model-as-a-service operator above US$50M worldwide gross revenue in any 12 months, or above 100M monthly active users, needs a separate written licence from Tencent; internal use is exempt (section 10.3). Read it before commercial self-hosting.

Worth knowing

Tencent's cost-tier MoE, weights published Nov 17, 2026 on Hugging Face (instruct, FP8 and base repositories, with ModelScope and GitCode mirrors): 220.3B total parameters by the model card's own table — 155.5B backbone + 63.5B Engram N-gram embedding tables + 0.43B SigLIP2 vision encoder + 0.85B DFly2 speculative-decoding drafter — with 11.3B text + 0.43B vision parameters active per token, 42 layers, 512 routed experts (16 active), native image input and a 1,048,576-token context. Memory by arithmetic: about 440 GB at BF16, about 220 GB in the published FP8 build. Serving recipes for vLLM and SGLang are published, including an official prebuilt vLLM image; Transformers support is still a pull request. All benchmarks are Tencent's own (WorkIndex 42.1; a blind test at 2.79/4 satisfaction against Claude Haiku 5.5's 2.81) — no independent evaluation exists yet.

What it needs

Full precision (FP16)
441 GB
the weights, uncompressed
8-bit
220 GB
the weights, compressed to 8-bit
4-bit
110 GB
the weights, compressed to 4-bit

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.

1.B300 Most recommended Runs it compressed288GB

Runs Hy4 flash compressed — the 8-bit build needs 220GB, which fits this card's 288GB. The compromise: slightly lower answer quality and a shorter conversation memory than the full-precision build.

Very costly for small jobs — best value only on big work.

Lowest verified: Hyperstack — $7.40/hr Checked Nov 21, 2026 Launch ↗
2.H200 Runs it compressed141GB

Runs Hy4 flash compressed — the 4-bit build needs 110GB, which fits this card's 141GB. The compromise: slightly lower answer quality and a shorter conversation memory than the full-precision build.

More card than most jobs need; the main trap is paying for it on small work.

Lowest verified: RunPod Community — $3.59/hr Checked Nov 21, 2026 Launch ↗
3.B200 Runs it compressed192GB

Runs Hy4 flash compressed — the 4-bit build needs 110GB, which fits this card's 192GB. The compromise: slightly lower answer quality and a shorter conversation memory than the full-precision build.

Frontier training money. If your job fits on anything cheaper, a B200 is budget burned for a spec sheet.

Lowest verified: RunPod Community — $5.98/hr Checked Nov 21, 2026 Launch ↗

At full precision, Hy4 flash needs more than one GPU — no single card holds its 441GB of weights plus room to work. Every card above runs a compressed build.

The cheapest verified way to run it

H200 at RunPod Community — $3.59/hr — runs it compressed — the 4-bit build.

Launch RunPod Community ↗  See the H200 →

Checked Nov 21, 2026

Select a GPU  See these GPUs on the prices board, ranked for this model.