Only with CPU offload

Qwen3.6 27B at Q4_K_M needs 16.7 GB but only 8.8 GB is addressable, so about 51% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 4.3 tokens per second — usable for batch work, painful for chat.

49% on GPU 8K context Q4_K_M · 15.6 GB Apache 2.0 Released 22 Apr 2026 Vision

The 24 GB coding pick of spring 2026 (77.2 SWE-bench Verified). Same shape as 3.8, one generation behind.

What hardware do I need for Qwen3.6 27B? →

The VRAM budget

weights 15.6 GB

Weights 15.6 GB KV cache @ 8K 0.50 GB Runtime overhead 0.6 GB Over budget 7.9 GB past 8.8 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
Q8_0	27.5 GB	28.6 GB	—	~1.8	−0.1% ppl	19.8 GB over
Q6_K	21.2 GB	22.3 GB	—	~2.6	−0.4% ppl	13.5 GB over
Q5_K_M	18.4 GB	19.5 GB	—	~3.2	−0.8% ppl	10.7 GB over
Q4_K_M	15.6 GB	16.7 GB	—	~4.3	−1.9% ppl	7.9 GB over
Q3_K_M	12.7 GB	13.8 GB	—	~6.5	−5.4% ppl	5.0 GB over
Q2_K	10.8 GB	11.9 GB	—	~9.7	−15% ppl	3.1 GB over

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

Ollama llama.cpp LM Studio vLLM

terminal

$ llama-server \
    -hf Qwen/Qwen3.6-27B:Q4_K_M \
    -c 8192 -ngl 31

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 15.6 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.

02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.

03If it falls back to CPU silently, drop the context first, then step down a quantisation.

See all models for this rig Compare with another model