Only with CPU offload

Qwen3.6 35B-A3B at Q4_K_M needs 20.9 GB but only 18.4 GB is addressable, so about 13% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 39 tokens per second — usable for batch work, painful for chat.

87% on GPU 8K context Q4_K_M · 20.2 GB Apache 2.0 Released 16 Apr 2026 Vision

Mixture of experts with ~3B active: the fastest serious model a 24 GB card runs, and the best MoE under 40B on agentic coding.

What hardware do I need for Qwen3.6 35B-A3B? →

Fits instead: Qwen3.6 27B (16.7 GB)

The VRAM budget

weights 20.2 GB

Weights 20.2 GB KV cache @ 8K 0.16 GB Runtime overhead 0.6 GB Over budget 2.5 GB past 18.4 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
Q8_0	35.5 GB	36.3 GB	—	~7.9	−0.1% ppl	17.9 GB over
Q6_K	27.4 GB	28.2 GB	—	~14	−0.4% ppl	9.8 GB over
Q5_K_M	23.7 GB	24.5 GB	—	~21	−0.8% ppl	6.1 GB over
Q4_K_M	20.2 GB	20.9 GB	—	~39	−1.9% ppl	2.5 GB over
Q3_K_M	16.3 GB	17.1 GB	74K	124	−5.4% ppl	Long context
Q2_K	14.0 GB	14.8 GB	194K	145	−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 10 of its 40 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/Qwen3.6-35B-A3B:Q4_K_M \
    -c 8192 -ngl 34

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

01Download is 20.2 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