GPUs tracked45Price history6 dayssince Sep 3, 2026Price points (24h)8,961 pointsBiggest 24h dropGeForce RTX 5090-4.3%Biggest 24h riseGeForce RTX 4080+20.0%Data updated

How to Deploy Hy3 Locally

Hy3 is a MoE model with 295B total parameters and 21B active per token. At Q4 it needs 168GB of VRAM to sit entirely on the GPU; with experts offloaded to system RAM it needs only 10GB of VRAM plus 160GB of RAM. The cheapest device that holds it is the Mac Studio M3 Ultra 512GB, with no used-price data yet, at roughly 30 tokens/s. In offload mode the cheapest card that runs it is the Tesla V100 16GB, at roughly 7 tokens/s.

Hy3 details

Launch dateJul 6, 2026
Hugging Face repotencent/Hy3
Revision
LicenseApache-2.0
ArchitectureMoE
Ollama tag
Min VRAM at Q4168 GB · Q4_K_M · 8K
Experts offloaded10 GB VRAM + 160 GB RAM
Parameters295B
MoE21B
Layers80
Hidden size4,096
KV heads8
Head dim128
Max context256K
VendorTencent

How much VRAM does Hy3 need for local deployment?

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

QuantizationWeightsGBTotal at 8KGBTotal at 32KGBTotal at 128KGB
1-bit71.1GGUF72.780.2110.2
2-bit107.2GGUF113.1120.6150.6
3-bit143.1GGUF144.9152.4182.4
4-bit182.2GGUF167.9175.4205.4
5-bit212.8GGUF202.5210240
8-bit317.7GGUF309.3316.8346.8
FP16577est.580.5588618

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.

Which GPUs can run Hy3?

Q4_K_M · 8K context · sorted by eBay used price
DeviceVRAMGBUsed priceEst. t/sSuggested context
Mac Studio M3 Ultra 512GB512No listings30128K

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.

Recommended GPUs for Hy3 with CPU/GPU offloading

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.

DeviceVRAMGBUsed priceRAM neededGBEst. t/s
Tesla V100 16GB16$2891606.5
Radeon RX 7800 XT16$5561606.1
Radeon RX 7900 XT20$6901606.3
Tesla V100 32GB32$7391606.5
Radeon RX 9070 XT16$7901606.1
Radeon RX 7900 XTX24$1,0001606.4
Instinct MI10032$1,0841606.5
GeForce RTX 408016$1,2601606.4
GeForce RTX 5070 Ti16$1,2891606.5
GeForce RTX 309024$1,5501606.6
GeForce RTX 508016$1,8001606.6
RTX PRO 4000 Blackwell24$2,9501606.4
GeForce RTX 409024$3,2001606.6
Radeon PRO W790048$3,4951606.4
RTX PRO 4500 Blackwell32$4,3001606.5
Instinct MI21064$5,0461606.6
MacBook Pro M4 Max 128GB128$5,4001606.1
GeForce RTX 509032$5,9251606.8
A100 40GB PCIe40$6,0001606.7
MacBook Pro M5 Max 128GB128$7,2501606.2
RTX PRO 6000 Blackwell96$16,9851606.8
GeForce RTX 5070121606.4
GeForce RTX 5060 Ti 16GB161606.1
Radeon RX 6950 XT161606.1
GeForce RTX 2080 Ti 22GB221606.3
Arc Pro B60241605.7
GeForce RTX 5090 D V2241606.7
Arc Pro B65321606
Arc Pro B70321606
GeForce RTX 4080 SUPER 32GB321606.4
GeForce RTX 5090 D321606.8
Instinct MI50 32GB321606.4
Radeon AI PRO R9700321606.1
Radeon PRO W7800321606.1
CMP 170HX 40GB401606.7
GeForce RTX 4090 48GB481606.6
RTX PRO 5000 Blackwell 48GB481606.7
CMP 170HX 64GB641606.7
RTX PRO 5000 Blackwell 72GB721606.7
RTX PRO 6000D841606.7
DGX Spark 128GB1281605.7
Mac Studio M4 Max 128GB1281606.1
Mac Studio M5 Max 128GB1281606.2
Ryzen AI Max+ 395 128GB1281605.2

Deploy Hy3 locally with Ollama, llama.cpp or vLLM

Ollama

The Ollama library has no official tag for it yet.

llama.cpp

llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.

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

Follow the deployment guide to prepare the required runtime and hardware.

# 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

FAQ

How much VRAM does Hy3 need?

At Q4_K_M with an 8K context Hy3 needs about 168GB of VRAM; 204GB leaves comfortable headroom.

What is the cheapest GPU that runs Hy3?

The Mac Studio M3 Ultra 512GB: 512GB of VRAM at roughly 30 tokens/s. We have no used-price data for it yet.

Can you run Hy3 with Ollama?

Not yet: the Ollama library has no official tag for it. You can load the GGUF with llama.cpp instead.