How much VRAM does Hy3 need?
Hy3 needs 167 GB for its weights at 4-bit, 170 GB with an 8K context and 177 GB with 32K. At 8-bit, 295 GB.
- Parameters
- 295B
- Active per token
- 21B
- Longest context
- 256K
- Released
- Jul 2026
- License
- Apache 2.0
- Strongest at
- Chat, Coding, Agents, Docs
Memory at each precision
| Precision | Weights | With 8K | With 32K | With 128K |
|---|---|---|---|---|
| 16-bit | 549 GB | 553 GB | 560 GB | 590 GB |
| 8-bit | 292 GB | 295 GB | 303 GB | 333 GB |
| 6-bit | 225 GB | 229 GB | 236 GB | 266 GB |
| 5-bit | 195 GB | 199 GB | 206 GB | 236 GB |
| 4-bit | 167 GB | 170 GB | 177 GB | 207 GB |
| 3-bit | 134 GB | 138 GB | 145 GB | 175 GB |
| 2-bit | 103 GB | 106 GB | 114 GB | 144 GB |
| 1.6-bit | 56.7 GB | 60.0 GB | 67.5 GB | 97.5 GB |
One conversation, 16-bit context cache, runtime overhead included.
What runs it
| Machine | Best fit | Context | Speed | First reply |
|---|---|---|---|---|
| RTX 3060 12 GB | Too big | |||
| RTX 4060 Ti 16 GB | Too big | |||
| RTX 3090 | Too big | |||
| RTX 4090 | Too big | |||
| RTX 5090 | Too big | |||
| 2× RTX 3090 | Too big | |||
| RTX PRO 6000 Blackwell | Too big | |||
| M4 Max 128 GB | Too big | |||
| M5 Max 128 GB | Too big | |||
| Ryzen AI Max+ 395 128 GB | Too big | |||
| DGX Spark 128 GB | 2-bit | 32K | 22 tok/s | 4.8 s |
| M3 Ultra 256 GB | 4-bit | 48K | 28 tok/s | 13 s |
| M5 Ultra 512 GB | 8-bit | 224K | 25 tok/s | 4.1 s |
| H100 SXM | Too big | |||
| DGX Station GB300 | 6-bit | 40K | 121 tok/s | 0.4 s |
| 8× H200 NVL | 8-bit | 256K | 87 tok/s | 0.2 s |
Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.
How to run it
- llama.cpp
llama-server -hf bartowski/Hy3-GGUF- vLLM
vllm serve tencent/Hy3- MLX (Mac)
mlx_lm.chat --model mlx-community/Hy3-OptiQ-2bit
Questions
How much VRAM does Hy3 need?
Hy3 needs 167 GB for its weights at 4-bit, 170 GB with an 8K context and 177 GB with 32K. At 8-bit, 295 GB.
Can Hy3 run on an RTX 4090?
Not on one card. The smallest machine here that runs it is the DGX Spark 128 GB.