Blog / Model guides
Qwen3.8-27B hardware requirements: what one cloud GPU needs
Qwen published Qwen3.8-27B on Hugging Face on 14 August 2026 under the Apache 2.0 license (huggingface.co/Qwen/Qwen3.8-27B). It is a dense 27-billion-parameter model that takes images as well as text. Simon Willison's note on the release puts the Q4_K_M build at about 17 GB on disk. SGLang reported 206.1 tokens per second single-stream decode on one RTX 5090 with NVFP4 on release day. A dense model needs exactly what its parameter count says at the precision you pick, which keeps the single-GPU arithmetic short.
Memory by precision
Weights only. The KV cache for your context length and concurrency comes on top.
| Precision | Weights in memory | Basis |
|---|---|---|
| bf16 | about 54 GB | Calculated: 27B parameters at 2 bytes each |
| 8-bit (fp8 or int8) | about 27 GB | Calculated: 1 byte per parameter |
| 4-bit (Q4_K_M, NVFP4, AWQ) | about 14 to 17 GB | 17 GB for Q4_K_M per Willison |
One 80 GB H100 holds bf16 with cache to spare. One 48 GB L40S, the GPU in an AWS g6e.xlarge, holds the 8-bit or 4-bit build with room for context. Threads on the model's Hugging Face discussion tab report 4-bit builds on 24 GB consumer cards; those are reports, not our results. For comparison, DeepSeek V4 Flash needs two H200s.
What the GPU costs
List prices checked on 22 August 2026.
| Shape | Per hour | Source |
|---|---|---|
| AWS g6e.xlarge, Stockholm (eu-north-1), on-demand | $1.974 | Spare Cores |
| AWS g6e.xlarge, Stockholm, spot | $0.604 | Spare Cores |
| AWS g6e.xlarge, Frankfurt (eu-central-1), on-demand | $2.327 | Spare Cores |
| RunPod L40S, community cloud | $0.79 | GetDeploying |
At the Stockholm on-demand rate a g6e.xlarge is about $1,440 a month around the clock (730 hours) and about $350 a month for business hours (176 hours) on a wake/sleep schedule, multiplied from the list price. The cost calculator does the same arithmetic for your own hours. Qwen3.8 models reason by default; Willison recommends setting reasoning effort to low for interactive use.
Questions people ask
Does Qwen3.8-27B fit on one L40S?
An 8-bit or 4-bit build fits on a single 48 GB L40S, the GPU in an AWS g6e.xlarge, with room for a working context window. The bf16 weights alone are about 54 GB, so bf16 does not fit on an L40S.
How much VRAM does Qwen3.8-27B need at bf16?
About 54 GB for the weights, derived from 27 billion parameters at 2 bytes each, plus KV cache for your context length and batch size. One 80 GB H100 or one 141 GB H200 holds it with cache to spare.
Is Qwen3.8-27B Apache 2.0?
Yes. Qwen published the weights on Hugging Face on 14 August 2026 under the Apache 2.0 license, which permits commercial use and self-hosting.
What is the cheapest cloud GPU to run Qwen3.8-27B on?
For a quantized build, a single L40S. On 22 August 2026 an L40S was listed at $0.79 per hour on RunPod community cloud and an AWS g6e.xlarge at $1.974 per hour on-demand in Stockholm ($0.604 spot). Those are list prices, not our measurements.