Model Spend Arena2ND ED.
299B · MoE · 262,144 ctx · 2026-08-14

Hy3

Runs on Data centre · Hugging Face ↗

vLLMSGLang

Coding index 58.8 (Artificial Analysis)

Tencent Hunyuan's ~299B MoE, open-weight — the same model whose hosted free window opened and closed on the leaderboard, here as a self-host option.

Sizes on disk

Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈11 GB at 32K, fp16).

QuantizationSize
Q3_K_M137.5 GB
IQ4_XS160.8 GB
Q4_0170.3 GB
Q4_K_S175.5 GB
Q4_K_M182.2 GB
Q4182.5 GB
Q4_1187.8 GB
Q5_K_M212.8 GB
Q2_K213.9 GB
Q6_K257.2 GB
Q8_0317.7 GB

Config tips

~182 GB at Q4 — multi-GPU. A GGUF exists, so llama.cpp with heavy offload is possible but slow; vLLM / SGLang tensor-parallel is the real route.