Qwen 3.5 27B VRAM Requirements
Qwen 3.5 27B needs about 16.9 GB of VRAM at Q4_K_M with an 8K context: 15.2 GB for weights, 0.6 GB for the KV cache and 1.1 GB of runtime overhead. At Q8_0 it needs 28.3 GB and at full FP16 precision 51.8 GB. The smallest GPU that runs it fully is the Radeon RX 7900 XT 20GB.
16.9 GB VRAM
Qwen 3.5 27B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 15.2 GB
- KV cache 0.6 GB
- Runtime overhead 1.1 GB
Which GPUs can run it
- RTX 3060 12GB67% on GPU~5.5 tok/s
- RTX 4060 Ti 16GB89% on GPU~8.1 tok/s
- RTX 3090 24GBFits~36 tok/s
- Radeon RX 7900 XTX 24GBFits~31 tok/s
- RTX 4090 24GBFits~38 tok/s
- RTX 5090 32GBFits~67 tok/s
- RTX 6000 Ada 48GBFits~36 tok/s
- A100 80GBFits~75 tok/s
- H100 80GB SXMFits~120 tok/s
- Ryzen AI Max+ 395 128GBFits~9.1 tok/s
- Apple M4 Max 128GBFits~21 tok/s
- Apple M5 Max 128GBFits~23 tok/s
- RTX PRO 6000 Blackwell 96GBFits~67 tok/s
- NVIDIA DGX Spark 128GBFits~9.7 tok/s
- Apple M3 Ultra 512GBFits~31 tok/s
Qwen 3.5 27B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 15.2 GB | 16.9 GB | 18.4 GB | 20 GB |
| Q5_K_M | 5.69 | 17.8 GB | 19.5 GB | 21.0 GB | 24 GB |
| Q6_K | 6.59 | 20.6 GB | 22.3 GB | 23.8 GB | 24 GB |
| Q8_0 | 8.50 | 26.6 GB | 28.3 GB | 29.8 GB | 32 GB |
| FP16 / BF16 | 16.00 | 50.1 GB | 51.8 GB | 53.3 GB | 64 GB |
| Q3_K_M | 3.91 | 12.2 GB | 13.9 GB | 15.4 GB | 16 GB |
| Q2_K | 3.00 | 9.4 GB | 11.1 GB | 12.6 GB | 12 GB |
VRAM by context length
At Q4_K_M. Hybrid attention: only 16 of 64 layers use full attention and keep a KV cache (4 KV heads × 256 dimensions). The other 48 use linear attention (Gated DeltaNet) with a fixed state of about 150 MB that does not grow with context.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.4 GB | 0.3 GB | 16.6 GB | 16.5 GB |
| 8K tokens | 0.6 GB | 0.4 GB | 16.9 GB | 16.7 GB |
| 32K tokens | 2.1 GB | 1.2 GB | 18.4 GB | 17.4 GB |
| 128K tokens | 8.1 GB | 4.4 GB | 24.4 GB | 20.6 GB |
| 256K tokens | 16.1 GB | 8.6 GB | 32.4 GB | 24.9 GB |
GPUs that can run Qwen 3.5 27B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| Radeon RX 7900 XT 20GB | 20.0 GB | Fits | Q4_K_M | 38K | ~26 tok/s |
| Apple M5 32GB | 21.3 GB | Fits | Q5_K_M | 58K | ~5.9 tok/s |
| NVIDIA L4 24GB | 24.0 GB | Fits | Q6_K | 98K | ~12 tok/s |
| Tesla P40 24GB | 24.0 GB | Fits | Q6_K | 98K | ~11 tok/s |
| Apple M4 Max 36GB | 24.0 GB | Fits | Q6_K | 98K | ~16 tok/s |
| Intel Arc Pro B60 24GB | 24.0 GB | Fits | Q6_K | 98K | ~14 tok/s |
| RTX 3090 24GB | 24.0 GB | Fits | Q6_K | 98K | ~36 tok/s |
| Radeon RX 7900 XTX 24GB | 24.0 GB | Fits | Q6_K | 98K | ~31 tok/s |
| RTX 3090 Ti 24GB | 24.0 GB | Fits | Q6_K | 98K | ~38 tok/s |
| RTX 4090 24GB | 24.0 GB | Fits | Q6_K | 98K | ~38 tok/s |
| Radeon AI PRO R9700 32GB | 32.0 GB | Fits | Q8_0 | 219K | ~21 tok/s |
| RTX 5090 32GB | 32.0 GB | Fits | Q8_0 | 219K | ~67 tok/s |
| Apple M4 Pro 48GB | 36.0 GB | Fits | Q8_0 | 256K | ~11 tok/s |
| A100 40GB | 40.0 GB | Fits | Q8_0 | 256K | ~58 tok/s |
| RTX 4060 Ti 16GB | 16.0 GB | 89% on GPU | IQ4_XS | – | ~8.1 tok/s |
| RTX 4070 Ti Super 16GB | 16.0 GB | 89% on GPU | IQ4_XS | – | ~13 tok/s |
| RTX 4080 16GB | 16.0 GB | 89% on GPU | IQ4_XS | – | ~13 tok/s |
| RTX 4080 Super 16GB | 16.0 GB | 89% on GPU | IQ4_XS | – | ~13 tok/s |
| RTX 5060 Ti 16GB | 16.0 GB | 89% on GPU | IQ4_XS | – | ~10 tok/s |
| RTX 5070 Ti 16GB | 16.0 GB | 89% on GPU | IQ4_XS | – | ~14 tok/s |
About Qwen 3.5 27B
The first dense Qwen with hybrid attention: 16 gated full-attention layers and 48 Gated DeltaNet layers. That makes long context unusually cheap for a model this size, and it is the reason a 24 GB card can hold it at Q4_K_M with a 64K+ context. Later releases (Qwen 3.6 27B and 3.8 27B) keep the exact same architecture, so memory figures are identical.
| Spec | Value |
|---|---|
| Parameters | 26.9B |
| Layers | 64 |
| Hidden size | 5120 |
| Attention heads / KV heads | 24 / 4 |
| Head dimension | 256 |
| Vocabulary | 248,320 |
| Max context (used here) | 256K tokens |
| Linear-attention layers | 48 of 64 (Gated DeltaNet, fixed 150 MB state) |
| KV cache per token (FP16) | 64 KB |
| Released | 2026-02 |
| License | Apache 2.0 |
Qwen 3.5 27B VRAM: common questions
Does Qwen 3.5 27B fit on a 24GB GPU?
Yes. At Q4_K_M the weights are about 15 GB, and the KV cache only grows on 16 of 64 layers, so a 24 GB card has room for a long context.
Why does my runtime not support Qwen 3.5?
Gated DeltaNet layers need runtime support. Use a recent llama.cpp, Ollama, LM Studio or vLLM build; older versions fail to load the model.
How much VRAM does Qwen 3.5 27B need?
About 16.9 GB at Q4_K_M with an 8K context, 28.3 GB at Q8_0 and 51.8 GB at FP16. Longer contexts add 0.06 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Qwen 3.5 27B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the Radeon RX 7900 XT 20GB (20.0 GB usable).