Yes — with 29.7 GB to spare
Qwen3.5-27B at Q4_K_M fits your L40S entirely on the GPU at 8K context, at an estimated 33 tokens per second. There is room for its full 256K window.
Sized from Qwen/Qwen3.5-27B's config.json — not a hand-verified catalogue entry.
Read from config.json
Parameter count is safetensors index; the active count for mixture-of-experts models is computed from expert sizes and may differ from the model card by a few percent. Pipeline: image-text-to-text · 1,888,406 downloads.
GGUF conversions found
- unsloth/Qwen3.5-27B-GGUF 180,523 downloads
- bartowski/Qwen_Qwen3.5-27B-GGUF 163,457 downloads
- Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF 26,639 downloads
- wcn123/Qwen3.5-27B-WebNovel-Writer-zh-GGUF 17,792 downloads
- lmstudio-community/Qwen3.5-27B-GGUF 10,244 downloads
- unsloth/Qwen3.5-27B-MTP-GGUF 6,093 downloads
The llama.cpp command above uses the original repo id; point -hf at one of these if the original has no GGUF files.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 27.5 GB | 28.6 GB | 256K | 19 | −0.1% ppl | Long context |
| Q6_K | 21.2 GB | 22.3 GB | 256K | 25 | −0.4% ppl | Long context |
| Q5_K_M | 18.3 GB | 19.4 GB | 256K | 29 | −0.8% ppl | Long context |
| Q4_K_M | 15.6 GB | 16.7 GB | 256K | 33 | −1.9% ppl | Recommended |
| Q3_K_M | 12.6 GB | 13.7 GB | 256K | 41 | −5.4% ppl | Long context |
| Q2_K | 10.8 GB | 11.9 GB | 256K | 48 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 16 of its 64 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ llama-server \
-hf Qwen/Qwen3.5-27B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.