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 card that runs it today is the Mac Studio M3 Ultra 512GB at about $19,315, 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
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.
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| 1-bit | 71.1GGUF | 72.7 | 80.2 | 110.2 |
| 2-bit | 107.2GGUF | 113.1 | 120.6 | 150.6 |
| 3-bit | 143.1GGUF | 144.9 | 152.4 | 182.4 |
| 4-bit | 182.2GGUF | 167.9 | 175.4 | 205.4 |
| 5-bit | 212.8GGUF | 202.5 | 210 | 240 |
| 8-bit | 317.7GGUF | 309.3 | 316.8 | 346.8 |
| FP16 | 577est. | 580.5 | 588 | 618 |
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?
Xianyu
| Device | VRAMGB | Price | Est. t/s | Suggested context |
|---|---|---|---|---|
| Mac Studio M3 Ultra 512GB | 512 | $19,3156 listings | 30 | 128K |
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
Xianyu
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.
Deploy Hy3 locally with Ollama, llama.cpp or vLLM
The Ollama library has no official tag for it yet.
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
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, currently about $19,315, at roughly 30 tokens/s.
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.