Qwen3.6-35B-A3B VRAM and GPU requirements (256K context)

The short answer

We measured about 102 tok/s decode on one Radeon AI PRO R9700 (llama.cpp Vulkan, short prompt) on the local desktop and run UD-Q5_K_XL daily at the full 262K context. Cards with 24 GB or less still run it by keeping some MoE experts in system RAM, but slower. Two R9700s are not needed for this model: “35B on two cards” means one full copy per card.

Every Radeon we cover, at 256K

Sizes are Linux GiB, shown as GB. Fit is a calculator estimate at 262,144 tokens with f16 KV. Q4 and up. Q3 is a labeled estimate on the can-run leaves, not a house pick.

How fast is offload?

One 16 GB RX 9070 XT, Q4_K_M, ROCm, DDR4-3200: 31.8 to 43.9 tok/s decode as --n-cpu-moe goes 40 to 16 (community report; third-party). One R9700 with everything in VRAM on the local desktop: 117.7 tok/s Q4_K_M, 102.4 tok/s UD-Q5_K_XL. Different hosts, so treat the gap as an estimate.

• 35B on RX 9070 XT
• Out of VRAM
• MoE

The honest pick

Already have 32 GB or more: buy nothing. 16–24 GB and at least 64 GB RAM: try offload first. 16–24 GB and under 64 GB RAM: a DDR5 kit is the cheaper fix. Want the full-speed, full-context Q5: one 32 GB R9700. Two R9700s are not needed for this model; the second card is for Flash-class models. 96 GB DDR5 is the house RAM size for most builds.

• R9700 listings
• 96 GB DDR5 kit
• One vs two GPUs

Launch lines (Q4+, 256K)

32 GB cards use the house UD-Q5_K_XL file. 16–24 GB cards start --n-cpu-moe at 24 and lower it until the load fits. llama.cpp first.

./build-vk/bin/llama-server \
  -m /path/to/nvme/Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf \
  -ngl 99 -c 262144 --cache-type-k f16 --cache-type-v f16 --jinja \
  --host 127.0.0.1 --port 8080

# 16–24 GB expert offload
./build-vk/bin/llama-server \
  -m /path/to/nvme/Qwen3.6-35B-A3B-Q4_K_M.gguf \
  -ngl 99 --n-cpu-moe 24 -c 262144 -t <physical-cores> \
  --cache-type-k f16 --cache-type-v f16 --jinja \
  --host 127.0.0.1 --port 8080

• House 35B recipe
• llama.cpp

Where to look

If you want the full-speed Q5 at 256K and you do not already have a 32 GB card, check current R9700 listings. We do not print a street price here. AMD’s launch MSRP is $1,299. 96 GB DDR5 is the house RAM pick for most builds.

Check R9700 listings

FAQ

How much VRAM does Qwen3.6-35B-A3B need?

At 256K context with f16 KV: about 32.7 GB for Q6_K, 30.7 GB for UD-Q5_K_XL, and 25.6 GB for Q4_K_M (calculator estimates). A 32 GB Radeon holds the Q5 fully. llama.cpp first.

35B requirements · Calculator

Can a 16 GB GPU run Qwen3.6-35B?

Yes, with expert offload and enough system RAM. Use Q4_K_M and --n-cpu-moe. It is slower than a 32 GB card that holds Q5 in VRAM. Third-party 9070 XT reports land around 32–44 tok/s. Not measured on the local desktop for that card.

35B on RX 9060 XT · Out of VRAM

Do I need two GPUs for Qwen3.6-35B?

No. One 32 GB card holds the Q5 at 256K. “35B on two cards” on this site means one full copy per card, not a split. A split was slower than one card at every depth we measured on the local desktop.

35B on two cards · One vs two GPUs

Sources

• S-CALC — AMDai calculator — 35B Q6/Q5/Q4 totals at 256K — our lab · checked 2026-11-02 · site calculator
• S-LAB — Local desktop 35B llama.cpp Vulkan cards — our lab · checked 2026-11-02 · Measured on the local desktop
• S-HF — Qwen3.6-35B-A3B model card — third-party · checked 2026-11-02 · read 2026-11-02
• S-GGUF — Unsloth Qwen3.6-35B-A3B GGUF listing — third-party · checked 2026-11-02 · read 2026-11-02