How much VRAM does Qwen3.5 397B A17B need?
Qwen3.5 397B A17B needs 228 GB for its weights at 4-bit, 229 GB with an 8K context and 230 GB with 32K. At 8-bit, 401 GB.
- Parameters
- 403B
- Active per token
- 17B
- Longest context
- 256K
- Released
- Feb 2026
- License
- Apache 2.0
- Strongest at
- Agents, Docs
Memory at each precision
| Precision | Weights | With 8K | With 32K | With 128K |
|---|---|---|---|---|
| 16-bit | 751 GB | 753 GB | 754 GB | 756 GB |
| 8-bit | 399 GB | 401 GB | 401 GB | 404 GB |
| 6-bit | 308 GB | 310 GB | 310 GB | 313 GB |
| 5-bit | 267 GB | 269 GB | 269 GB | 272 GB |
| 4-bit | 228 GB | 229 GB | 230 GB | 233 GB |
| 3-bit | 184 GB | 185 GB | 186 GB | 189 GB |
| 2-bit | 141 GB | 142 GB | 143 GB | 146 GB |
| 1.6-bit | 77.5 GB | 78.9 GB | 79.7 GB | 82.5 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 | 1.6-bit | 256K | 153 tok/s | 0.8 s |
| M4 Max 128 GB | 1.6-bit | 256K | 53 tok/s | 18 s |
| M5 Max 128 GB | 1.6-bit | 256K | 57 tok/s | 5.1 s |
| Ryzen AI Max+ 395 128 GB | 1.6-bit | 256K | 34 tok/s | 10 s |
| DGX Spark 128 GB | 1.6-bit | 256K | 44 tok/s | 3.6 s |
| M3 Ultra 256 GB | 2-bit | 256K | 48 tok/s | 9.9 s |
| M5 Ultra 512 GB | 6-bit | 256K | 38 tok/s | 3.1 s |
| H100 SXM | Too big | |||
| DGX Station GB300 | 4-bit | 256K | 175 tok/s | 0.3 s |
| 8× H200 NVL | 8-bit | 256K | 111 tok/s | 0.1 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 unsloth/Qwen3.5-397B-A17B-GGUF- vLLM
vllm serve Qwen/Qwen3.5-397B-A17B- MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen3.5-397B-A17B-4bit
Questions
How much VRAM does Qwen3.5 397B A17B need?
Qwen3.5 397B A17B needs 228 GB for its weights at 4-bit, 229 GB with an 8K context and 230 GB with 32K. At 8-bit, 401 GB.
Can Qwen3.5 397B A17B run on an RTX 4090?
Not on one card. The smallest machine here that runs it is the RTX PRO 6000 Blackwell.