How to Deploy Hy3 Locally

MoE: all 295B parameters must stay resident in memory, while speed is set by the 21B active per token.

Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 256K limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.

Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.

Expert weights live in system RAM (needs at least 160 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 10 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.

Device	VRAMGB	Price	RAM neededGB	Est. t/s
Tesla V100 16GB	16	$659	160	6.5
Radeon RX 9070 XT	16	$884	160	6.1
GeForce RTX 5060 Ti 16GB	16	$924	160	6.1
GeForce RTX 5070	12	$956	160	6.4
Radeon RX 7800 XT	16	$966	160	6.1
Tesla V100 32GB	32	$1,141	160	6.5
Arc Pro B60	24	$1,164	160	5.7
Radeon RX 7900 XT	20	$1,244	160	6.3
Radeon RX 6950 XT	16	$1,404	160	6.1
GeForce RTX 5070 Ti	16	$1,576	160	6.5
Radeon RX 7900 XTX	24	$1,738	160	6.4
GeForce RTX 5080	16	$1,868	160	6.6
Arc Pro B70	32	$2,074	160	6
GeForce RTX 4080	16	$2,187	160	6.4
Radeon AI PRO R9700	32	$2,244	160	6.1
GeForce RTX 3090	24	$2,520	160	6.6
Ryzen AI Max+ 395 128GB	128	$3,913	160	5.2
RTX PRO 4000 Blackwell	24	$3,956	160	6.4
GeForce RTX 4090	24	$4,905	160	6.6
GeForce RTX 5090	32	$7,489	160	6.8
DGX Spark 128GB	128	$8,223	160	5.7
GeForce RTX 2080 Ti 22GB	22	—	160	6.3
GeForce RTX 5090 D V2	24	—	160	6.7
Arc Pro B65	32	—	160	6
GeForce RTX 4080 SUPER 32GB	32	—	160	6.4
GeForce RTX 5090 D	32	—	160	6.8
Instinct MI100	32	—	160	6.5
Instinct MI50 32GB	32	—	160	6.4
Radeon PRO W7800	32	—	160	6.1
RTX PRO 4500 Blackwell	32	—	160	6.5
A100 40GB PCIe	40	—	160	6.7
CMP 170HX 40GB	40	—	160	6.7
GeForce RTX 4090 48GB	48	—	160	6.6
Radeon PRO W7900	48	—	160	6.4
RTX PRO 5000 Blackwell 48GB	48	—	160	6.7
CMP 170HX 64GB	64	—	160	6.7
Instinct MI210	64	—	160	6.6
RTX PRO 5000 Blackwell 72GB	72	—	160	6.7
RTX PRO 6000D	84	—	160	6.7
RTX PRO 6000 Blackwell	96	—	160	6.8
Mac Studio M4 Max 128GB	128	—	160	6.1
Mac Studio M5 Max 128GB	128	—	160	6.2
MacBook Pro M4 Max 128GB	128	—	160	6.1
MacBook Pro M5 Max 128GB	128	—	160	6.2

llama-server -hf bartowski/Hy3-GGUF:Q4_K_M --jinja -c 8192

# Switch to trtllm backend to work-around mnnvl workspace size issue.
export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
vllm serve tencent/Hy3 \
  --tensor-parallel-size 8 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 2 \
  --tool-call-parser hy_v3 \
  --reasoning-parser hy_v3 \
  --enable-auto-tool-choice \
  --port 8000 \
  --served-model-name hy3

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology