Configs › Qwen3.6-35B-A3B › Qwen3.6-35B-A3B on Strix Halo: 127.1 tok/s, 29 measured configs

Qwen3.6-35B-A3B on Strix Halo: 127.1 tok/s, 29 measured configs

Measured Qwen3.6-35B-A3B speeds on Strix Halo: fastest 127.1 tok/s single-stream (UD-IQ4_XS, llama.cpp/vulkan (speculative decode, mtp-2)), 29 configs with quant, backend and context, each linked to its source, snapshot 2026-10-29.

Snapshot 2026-10-29 · 413 configs on AI boxes and 288 on GPU cards · 121 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.

Fastest measured single-stream run of Qwen3.6-35B-A3B on Strix Halo: 127.1 tok/s (UD-IQ4_XS, llama.cpp/vulkan (speculative decode, mtp-2)), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.6-35B-A3B on Strix Halo

Best quant on Strix Halo

Fastest plain-decoding run per quant, versus the fastest quant.

• Q4_0: 81.3 tok/s, fastest (llama.cpp/vulkan)
• Q4_K_M: 63.8 tok/s, 22% lower (llama.cpp/vulkan)
• UD-Q4_K_XL: 62.8 tok/s, 23% lower (llama.cpp/vulkan-radv)
• UD-Q8_K_XL: 49.2 tok/s, 39% lower (llama.cpp/vulkan-radv)

Backends on Strix Halo

Fastest plain-decoding run per backend. Quants can differ between rows.

• llama.cpp/vulkan: 81.3 tok/s (Q4_0)
• llama.cpp/vulkan-radv: 62.8 tok/s (UD-Q4_K_XL)
• llama.cpp/vulkan-radv-performance: 62.2 tok/s (UD-Q4_K_XL)
• ollama/vulkan: 60.6 tok/s (Q4_K_M)
• llama.cpp/vulkan-radv-mtp: 54.1 tok/s (UD-Q4_K_XL)
• llama.cpp/rocm-7.14: 53.5 tok/s (UD-Q4_K_XL)
• llama.cpp/rocm-7.14-pr26592: 53 tok/s (UD-Q4_K_XL)
• llama.cpp/rocm: 52.7 tok/s (Q4_K_M)
• llama.cpp/rocm-7.2.4: 51.7 tok/s (UD-Q4_K_XL)
• llama.cpp/rocm-7.2.3-mtp: 46 tok/s (UD-Q4_K_XL)

Qwen3.6-35B-A3B on other boxes

• Mac Max: 151.1 tok/s best of 3 configs
• Mac Ultra: 109.1 tok/s best of 1 config
• DGX Spark: 77.4 tok/s best of 22 configs

Questions and answers

How fast is Qwen3.6-35B-A3B on Strix Halo?

The fastest of 29 measured configs is 127.1 tok/s single-stream (UD-IQ4_XS, llama.cpp/vulkan (speculative decode, mtp-2)). The slowest is 39 tok/s.

Which quant of Qwen3.6-35B-A3B is fastest on Strix Halo?

Q4_0 at 81.3 tok/s, ahead of Q4_K_M at 63.8, UD-Q4_K_XL at 62.8, UD-Q8_K_XL at 49.2. Plain decoding only.

Which backend is fastest for Qwen3.6-35B-A3B on Strix Halo?

llama.cpp/vulkan at 81.3 tok/s, then llama.cpp/vulkan-radv at 62.8, llama.cpp/vulkan-radv-performance at 62.2, ollama/vulkan at 60.6, llama.cpp/vulkan-radv-mtp at 54.1, llama.cpp/rocm-7.14 at 53.5, llama.cpp/rocm-7.14-pr26592 at 53, llama.cpp/rocm at 52.7, llama.cpp/rocm-7.2.4 at 51.7, llama.cpp/rocm-7.2.3-mtp at 46. Each backend figure is its best plain-decoding row and may use a different quant.

How does Strix Halo compare with other boxes for Qwen3.6-35B-A3B?

Strix Halo best 127.1 tok/s. Other boxes, best measured: Mac Max 151.1, Mac Ultra 109.1, DGX Spark 77.4. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: Qwen3.6-35B-A3B on every box · Strix Halo hardware page (from S$4,329 in Singapore today).

Last updated: snapshot 2026-10-29; newest source check 2026-10-29.

More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.