Hunyuan Hy3 GPU requirements: 149 GB at 4-bit, 598 GB at BF16
Hunyuan Hy3 needs 598 GB of GPU memory for its weights at BF16, 299 GB at FP8 and 149 GB at 4-bit, plus 10.7 GB of KV cache for a 32K context. The cheapest node that runs it here is 4× A40 at 4-bit, for $0.24 an hour on spot.
Weights plus a 32K 16-bit KV cache, with 10% of the node's memory kept free. Billed per minute.
- Parameters
- 295B21B active per token
- Weights, 4-bit
- 149 GBBF16: 598 GB
- KV cache, 32K
- 10.7 GBOne sequence, 16-bit cache
- Context window
- 256KNative
How much VRAM Hunyuan Hy3 needs, format by format.
Weights are parameters × bytes per parameter. Each context column adds the KV cache of one sequence of that length.
| Format | Weights | + 8Kcontext | + 32Kcontext | + 128Kcontext | + 256Kcontext | Cheapest node, 32K |
|---|---|---|---|---|---|---|
| BF16Published weights, fine-tuning | 598 GB | 600 GB | 608 GB | 641 GB | 683 GB | 8× RTX PRO 6000 Blackwell · $1.52/h |
| FP8Hopper, Ada, Blackwell, MI300 | 299 GB | 301 GB | 310 GB | 342 GB | 385 GB | 8× L40 · $0.72/h |
| 4-bitAWQ, GPTQ, GGUF Q4 | 149 GB | 152 GB | 160 GB | 192 GB | 235 GB | 4× A40 · $0.24/h |
The serving engine adds its own buffers, and an 8-bit KV cache halves the context columns. We keep 10% of a node's memory free when we say a model fits. Speed is a separate question: it follows memory bandwidth and, for a mixture of experts, the active parameters.
GPUs that run Hunyuan Hy3, cheapest first.
For each format, the smallest node of each GPU model that holds the weights and a 32K context, sorted by spot price. One row per GPU model; larger nodes of the same GPU work too.
| Format | GPU | GPU memory | Spot | On-demand | 730 h spot | Action |
|---|---|---|---|---|---|---|
| 4-bit | 4× A40Ampere · 160 GB needed | 192 GB | $0.24 | $0.36 | $175.20 | Deploy |
| 4-bit | 4× L40Ada Lovelace · 160 GB needed | 192 GB | $0.36 | $0.76 | $262.80 | Deploy |
| 4-bit | 4× RTX 6000 AdaAda Lovelace · 160 GB needed | 192 GB | $0.36 | $0.76 | $262.80 | Deploy |
| 4-bit | 4× RTX A6000Ampere · 160 GB needed | 192 GB | $0.36 | $0.68 | $262.80 | Deploy |
| FP8 | 8× L40Ada Lovelace · 310 GB needed | 384 GB | $0.72 | $1.52 | $525.60 | Deploy |
| FP8 | 4× RTX PRO 6000 BlackwellBlackwell · 310 GB needed | 384 GB | $0.76 | $1.56 | $554.80 | Deploy |
| FP8 | 8× L40SAda Lovelace · 310 GB needed | 384 GB | $1.20 | $2.00 | $876.00 | Deploy |
| FP8 | 4× H100 NVLHopper · 310 GB needed | 376 GB | $1.96 | $3.16 | $1,430.80 | Deploy |
| BF16 | 8× RTX PRO 6000 BlackwellBlackwell · 608 GB needed | 768 GB | $1.52 | $3.12 | $1,109.60 | Deploy |
| BF16 | 8× H100 NVLHopper · 608 GB needed | 752 GB | $3.92 | $6.32 | $2,861.60 | Deploy |
| BF16 | 4× MI300XCDNA 3 · 608 GB needed | 768 GB | $3.96 | $5.80 | $2,890.80 | Deploy |
| BF16 | 4× MI325XCDNA 3 · 608 GB needed | 1024 GB | $4.36 | $6.36 | $3,182.80 | Deploy |
Spot capacity can be reclaimed with a 2-minute notice; for a live endpoint, use on-demand. How spot works · serving a language model.
What a long context costs in memory.
Grouped-query attention, 8 KV heads of 128 across 80 layers.
| Context | KV cache, 1 sequence | 8 sequences |
|---|---|---|
| 8K tokens | 2.68 GB | 21.5 GB |
| 32K tokens | 10.7 GB | 85.9 GB |
| 128K tokens | 42.9 GB | 344 GB |
| 256K tokens | 85.9 GB | 687 GB |
Serving several users
KV cacheEvery concurrent request keeps its own cache, so memory for context grows with users × context length. A server batching 8 requests of 32K needs eight times the single-sequence figure — 85.9 GB here.
- Recommended node4× A40
- Weights at 4-bit149 GB
- Context that fits, 1 sequence70K
Paged attention (vLLM, SGLang) allocates the cache in blocks as requests grow, so a node holds more short requests than the worst case suggests.
Hunyuan Hy3, as published.
- Developer
- TencentTencent mixture of experts, 256K context
- Parameters
- 295B total, 21B active per token
- Architecture
- Mixture of experts, 80 layersGrouped-query attention, 8 KV heads of 128 across 80 layers
- Context window
- 262,144 tokensAs set in the published configuration
- Published format
- BF16, 598 GB
- Licence
- Apache 2.0
- Released
- 2 July 2026
- Source
- tencent/Hy3config.json read 26 October 2026
Serve it with vLLM
template: vLLMOn 4× A40, from the instance terminal:
$ vllm serve tencent/Hy3 \ --tensor-parallel-size 4 \ --max-model-len 32768
Very recent architectures need a recent vLLM; the model card gives the minimum version and any extra flags.
Hunyuan Hy3 on a GPU — the questions that come up
How much VRAM does Hunyuan Hy3 need?
For the weights alone: 598 GB at BF16, 299 GB at FP8 and 149 GB at 4-bit. Add 10.7 GB of KV cache for one 32K-token context, or 85.9 GB at the full 256K window, and keep about 10% of the memory free for activations.
Can Hunyuan Hy3 run on a single GPU?
One RTX 4090 (24 GB): no. One RTX 5090 (32 GB): no. One H100 SXM (80 GB): no. One H200 SXM (141 GB): no. One B200 SXM (180 GB): yes, at 4-bit. Those answers assume a 32K context and the widest format that still fits.
What is the cheapest way to run Hunyuan Hy3?
Here, 4× A40 at 4-bit: $0.24 per hour on spot and $0.36 on-demand, billed per minute — $175.20 for a 730-hour month on spot. That node has 192 GB of GPU memory for the 160 GB this setup needs.
Does Hunyuan Hy3 need less memory because only 21B parameters are active?
No. A mixture-of-experts model routes each token through a few experts, but every expert has to be resident in GPU memory, so the memory bill follows the 295B total. What the 21B active parameters change is the work per token: generation reads far less of the weights each step than a dense model of the same size would, which is why it is fast for its size.
Which format should I run it in?
BF16 keeps the published weights untouched and is the format for fine-tuning. FP8 halves the memory and runs in hardware on Hopper, Ada, Blackwell and MI300-class GPUs. 4-bit quantisation (AWQ, GPTQ, GGUF) quarters it, at a quality cost that is larger on small models than on big ones; check it against your own evaluation before relying on it.
How do I deploy it?
Pick the node in the table above and launch it with the vLLM template; the model downloads from Hugging Face (tencent/Hy3) on first start. Spread the weights over the node with --tensor-parallel-size equal to the GPU count, and set --max-model-len to the context you need. New architectures need a recent vLLM: check the model card for the minimum version.
Sizing another model? GPU requirements for 64 open models · serving a language model · fine-tuning.
Hunyuan Hy3 on 4× A40 for $0.24 an hour.
Launch the node with the vLLM template, point it at tencent/Hy3, stop it when you are done. Prepaid, billed by the minute.
Billed per minute from the moment the instance is reachable. Minimum credit $30, no subscription.