Can Strix Halo run Llama 3.3 70B?

Q6 / Q5 / Q4 plus a Q3 estimate

Weights are official GGUF listings when the model page has them. Totals include KV at f16. 1-GPU and 2-GPU columns use the same calculator as this share URL.

System RAM

When the file does not stay in 64 GB VRAM, the spill is system RAM. Shopper lines stay Minimum system RAM needed and RAM needed at full context for Q4_K_M on this card. Estimate unless a measured R9700 card overlays the kit.

llama.cpp launch

Vulkan first, then ROCm/HIP. Copy a house recipe when one exists. llama.cpp stays first. Strata / vLLM stay on /backends.

# llama.cpp Vulkan. House pin when a recipe exists; otherwise a shape only.
./build-vk/bin/llama-server \
  -m ~/models/Llama-3.3-70B-Instruct-Q4_K_M.gguf \
  -ngl 99 \
  -c 262144 \
  --cache-type-k f16 \
  --cache-type-v f16 \
  --host 127.0.0.1 --port 8080

# HIP shape only — house 35B cards are Vulkan. HIP decode is not measured yet.
./build-hip/bin/llama-server \
  -m ~/models/Llama-3.3-70B-Instruct-Q4_K_M.gguf \
  -ngl 99 \
  -c 262144 \
  --cache-type-k f16 \
  --cache-type-v f16 \
  --host 127.0.0.1 --port 8080

Same model, other cards

Llama 3.3 70B on the other published Radeon sheets.

• Llama 3.3 70B on R9700
• Llama 3.3 70B on PRO W7900
• Llama 3.3 70B on PRO W7800
• Llama 3.3 70B on PRO W7700
• Llama 3.3 70B on PRO W6800
• Llama 3.3 70B on RX 7900 XTX
• Llama 3.3 70B on RX 7900 XT
• Llama 3.3 70B on RX 9070 XT
• Llama 3.3 70B on RX 9070
• Llama 3.3 70B on RX 9070 GRE
• Llama 3.3 70B on RX 9060 XT
• Llama 3.3 70B on RX 7900 GRE
• Llama 3.3 70B on RX 7800 XT
• Llama 3.3 70B on RX 7700 XT
• Llama 3.3 70B on RX 7600 XT
• Llama 3.3 70B on RX 7600
• Llama 3.3 70B on RX 6800 XT
• Llama 3.3 70B on RX 6950 XT
• Llama 3.3 70B on RX 6700 XT

Other models on Strix Halo

• Qwen3.6-35B
• Qwen 3.8 Flash
• gpt-oss 20B
• Gemma 3 27B
• GLM 5.3 Flash
• Qwen3.6 27B
• Qwen3.8 27B

Check price

Existing /go/ routes for this GPU sheet. Buttons say Check price — we do not print a figure.

FAQ

Can Strix Halo run Llama 3.3 70B?

Fits with RAM offload on the calculator at the shopper Q4_K_M file and 256K. Q6 / Q5 / Q4 are listed. Q3_K_M is a labeled estimate. llama.cpp first.

This pair · Strix Halo · Llama 3.3 70B · Calculator

Which quant should I download?

Prefer the shopper Q4_K_M file when it still fits. Q6 / Q5 / Q4 are the listed shopper steps. Q3_K_M is a labeled estimate — not a house pick. Q1/Q2 stay off.

Quants · Q6 / Q5 / Q4 + Q3 estimate

How fast is it?

Measured decode is only the official local desktop 35B llama.cpp Vulkan sweep on one or two R9700s. Every other model × card stays not measured yet. Record a card with amdai-bench.

Contribute · amdai-bench · Benchmarks

Stack Lab

Measure the pair on your box. Doctor and planner are allowlisted. House daily / Flash profiles appear only for those house models.