How Much VRAM Does Qwen3.8 27B Need to Run Locally?
See devices that run itQwen3.8 27B is a dense model with 27.8B parameters. At Q4 it needs at least 19GB of VRAM, and 24GB is the comfortable amount for an 8K context. The cheapest device that holds it is the Radeon RX 7900 XT, with no used-price data yet, at roughly 23.3 tokens/s.
Qwen3.8 27B details
VRAM needed for Qwen3.8 27B by quantization and context
Total at 8K| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| Q4_K_M | 16.5GGUF | 18.5 | 24.5 | 48.5 |
| Q5_K_M | 19.8GGUF | 21.7 | 27.7 | 51.7 |
| Q8_0 | 29GGUF | 31.8 | 37.8 | 61.8 |
| FP16 | 54.7GGUF | 57.3 | 63.3 | 87.3 |
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.
Which GPUs can run Qwen3.8 27B?
Q4_K_M · 8K context · sorted by eBay used price| Device | VRAMGB | Used price | Est. t/s | Suggested context |
|---|---|---|---|---|
 Radeon RX 7900 XT | 20 | —No listings | 23.3 | 8K |
 Arc Pro B60 | 24 | —No listings | 12.1 | 8K |
 GeForce RTX 3090 | 24 | —No listings | 37.8 | 8K |
| GeForce RTX 4090 | 24 | —No listings | 40.6 | 8K |
| GeForce RTX 5090 D V2 | 24 | —No listings | 53.5 | 8K |
| Radeon RX 7900 XTX | 24 | —No listings | 27.9 | 8K |
| RTX PRO 4000 Blackwell | 24 | —No listings | 27.3 | 8K |
| Arc Pro B65 | 32 | —No listings | 16 | 32K |
| Arc Pro B70 | 32 | —No listings | 16 | 32K |
| GeForce RTX 4080 SUPER 32GB (modded) | 32 | —No listings | 29.9 | 32K |
| GeForce RTX 5090 | 32 | —No listings | 70.3 | 32K |
| GeForce RTX 5090 D | 32 | —No listings | 70.3 | 32K |
| Instinct MI100 | 32 | —No listings | 35.5 | 32K |
| Radeon AI PRO R9700 | 32 | —No listings | 18.7 | 32K |
| Radeon PRO W7800 | 32 | —No listings | 16.9 | 32K |
| RTX PRO 4500 Blackwell | 32 | —No listings | 36.2 | 32K |
| Tesla V100 32GB | 32 | —No listings | 36.3 | 32K |
| A100 40GB PCIe | 40 | —No listings | 61.5 | 32K |
| CMP 170HX 40GB (modded) | 40 | —No listings | 61.7 | 32K |
| GeForce RTX 4090 48GB (modded) | 48 | —No listings | 40.6 | 32K |
| Radeon PRO W7900 | 48 | —No listings | 25.1 | 32K |
| RTX PRO 5000 Blackwell 48GB | 48 | —No listings | 53.5 | 32K |
| CMP 170HX 64GB (modded) | 64 | —No listings | 59.1 | 128K |
| Instinct MI210 | 64 | —No listings | 46.8 | 128K |
| RTX PRO 5000 Blackwell 72GB | 72 | —No listings | 53.5 | 128K |
| RTX PRO 6000D | 84 | —No listings | 62 | 128K |
| RTX PRO 6000 Blackwell | 96 | —No listings | 70.3 | 128K |
| Mac Studio M4 Max 128GB | 128 | —No listings | 17.6 | 128K |
| Mac Studio M5 Max 128GB | 128 | —No listings | 19.7 | 128K |
| MacBook Pro M4 Max 128GB | 128 | —No listings | 17.6 | 128K |
| MacBook Pro M5 Max 128GB | 128 | —No listings | 19.7 | 128K |
| Mac Studio M3 Ultra 512GB | 512 | —No listings | 26.2 | 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.
Run Qwen3.8 27B 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 unsloth/Qwen3.8-27B-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve Qwen/Qwen3.8-27B
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Qwen3.8 27B need?
At Q4_K_M with an 8K context Qwen3.8 27B needs about 19GB of VRAM; 24GB leaves comfortable headroom.
What is the cheapest GPU that runs Qwen3.8 27B?
The Radeon RX 7900 XT: 20GB of VRAM at roughly 23.3 tokens/s. We have no used-price data for it yet.
Can you run Qwen3.8 27B with Ollama?
Not yet: the Ollama library has no official tag for it. You can load the GGUF with llama.cpp instead.