Configs › Qwen3.6-35B-A3B tok/s: Mac Max 151.1, Mac Ultra 109.1
Qwen3.6-35B-A3B tok/s: Mac Max 151.1, Mac Ultra 109.1
Measured Qwen3.6-35B-A3B speeds: 151.1 tok/s on Mac Max (MLX-4BIT, mlx); 109.1 tok/s on Mac Ultra. 56 configs, each linked to its source, snapshot 2026-10-16.
Snapshot 2026-10-16 · 394 configs on AI boxes and 423 on GPU cards · 123 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.
Quick answers
- How fast is Qwen3.6-35B-A3B on Mac Max?
- 151.1 tok/s single-stream decode (MLX-4BIT, mlx), the fastest of 3 measured configs on Apple M-series Max (MacBook Pro / Mac Studio Max), measured by emihiggins, source last checked 2026-09-29. source
- How fast is Qwen3.6-35B-A3B on Mac Ultra?
- 109.1 tok/s single-stream decode (MLX-4BIT, mlx), the only measured config on Apple Mac Studio (M-series Ultra), measured by theNetworkChuck, source last checked 2026-09-29. source
- How fast is Qwen3.6-35B-A3B on DGX Spark?
- 85 tok/s single-stream decode (FP8, metrale (speculative decode, mtp-1)), the fastest of 24 measured configs on NVIDIA DGX Spark (GB10), measured by Metrale, source last checked 2026-10-05. source
- How fast is Qwen3.6-35B-A3B on Strix Halo?
- 81.3 tok/s single-stream decode (Q4_0, llama.cpp/vulkan), the fastest of 28 measured configs on AMD Strix Halo (Ryzen AI Max+ 395), measured by hogeheer, source last checked 2026-10-14. source
- How fast is Qwen3.6-35B-A3B on RTX 5090?
- 210.9 tok/s single-stream decode (UD-IQ4_NL, llama.cpp/cuda), the fastest of 17 measured configs on NVIDIA RTX 5090 (32GB), measured by paulplee, source last checked 2026-10-16. source
- How fast is Qwen3.6-35B-A3B on H100 SXM?
- 188.1 tok/s single-stream decode (UD-Q4_K_M, llama.cpp/cuda), the fastest of 4 measured configs on NVIDIA H100 SXM (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.6-35B-A3B on H100 PCIe?
- 152.6 tok/s single-stream decode (UD-Q4_K_M, llama.cpp/cuda), the fastest of 4 measured configs on NVIDIA H100 PCIe (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.6-35B-A3B on A100 40GB?
- 124 tok/s single-stream decode (UD-Q4_K_M, llama.cpp/cuda), the fastest of 4 measured configs on NVIDIA A100 40GB, measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.6-35B-A3B on A10?
- 95.4 tok/s single-stream decode (UD-Q4_K_M, llama.cpp/cuda), the fastest of 2 measured configs on NVIDIA A10 (24GB), measured by sypherin, source last checked 2026-10-14. source
Measured configs for Qwen3.6-35B-A3B on AI boxes
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.6-35B-A3B | Mac Max | MLX-4BIT | mlx | 151.1 tok/s (prefill 3333) | 512 | Extracted, source-linked | source |
| Qwen3.6-35B-A3B | Mac Max | Q4_K_M | llama.cpp | 124.3 tok/s (prefill 3343) | 512 | Extracted, source-linked | source |
| Qwen3.6-35B-A3B | Mac Ultra | MLX-4BIT | mlx | 109.1 tok/s (prefill 6012) | 32000 | Extracted, source-linked | source |
| Qwen3.6-35B-A3B | DGX Spark | FP8 | metrale (speculative decode, mtp-1) | 85 tok/s | 2048 | Extracted, source-linked | source |
| Qwen3.6-35B-A3B | Strix Halo | Q4_0 | llama.cpp/vulkan | 81.3 tok/s (prefill 1243.5) | 512 | Structured table | source |
| Qwen3.6-35B-A3B | DGX Spark | NVFP4 | vllm | 77.4 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan (speculative decode, mtp-3) | 74.6 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan (speculative decode, mtp-2) | 72.8 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/rocm (speculative decode, mtp-3) | 68.3 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | DGX Spark | UD-IQ3_XXS | llama.cpp/cuda | 66.3 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | DGX Spark | UD-IQ3_S | llama.cpp/cuda | 65.6 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan-radv-mtp (speculative decode, mtp-3) | 65.6 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/rocm (speculative decode, mtp-2) | 64.5 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan-radv-mtp (speculative decode, mtp-2) | 64.4 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | DGX Spark | UD-Q3_K_XL | llama.cpp/cuda | 64.1 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | DGX Spark | UD-Q3_K_M | llama.cpp/cuda | 64 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | Strix Halo | Q4_K_M | llama.cpp/vulkan | 63.8 tok/s (prefill 1064) | 512 | Structured table | source |
| Qwen3.6-35B-A3B | DGX Spark | UD-IQ4_NL | llama.cpp/cuda | 63.8 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | DGX Spark | UD-IQ4_XS | llama.cpp/cuda | 63.3 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | DGX Spark | UD-IQ4_NL_XL | llama.cpp/cuda | 63.1 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan-radv | 62.8 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | DGX Spark | Q4_K_XL | llama.cpp | 62.7 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan-radv-performance | 62.2 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q4_K_XL | llama.cpp/rocm-7.2.3-mtp (speculative decode, mtp-3) | 62.1 tok/s | Structured table | source | |
| Qwen3.6-35B-A3B | DGX Spark | UD-Q4_K_S | llama.cpp/cuda | 62.1 tok/s | 8192 | Structured table | source |
Qwen3.6-35B-A3B on GPU cards, ranked on their own
GPU cards run the model from their own VRAM: very fast on what fits, but most cannot load what a 128 GB AI box can. Ranked on their own, never against the boxes.
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.6-35B-A3B | RTX 5090 | UD-IQ4_NL | llama.cpp/cuda | 210.9 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q3_K_M | llama.cpp/cuda | 209.5 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-IQ4_NL_XL | llama.cpp/cuda | 208.8 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q3_K_XL | llama.cpp/cuda | 208.2 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-IQ4_XS | llama.cpp/cuda | 208.1 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q4_K_S | llama.cpp/cuda | 207.7 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-IQ3_XXS | llama.cpp/cuda | 204.7 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | MXFP4 | llama.cpp/cuda | 204.5 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q4_K_XL | llama.cpp/cuda | 203.9 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-IQ3_S | llama.cpp/cuda | 203.2 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q5_K_M | llama.cpp/cuda | 202.9 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q5_K_XL | llama.cpp/cuda | 201.8 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q4_K_M | llama.cpp/cuda | 201.6 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q3_K_S | llama.cpp/cuda | 199.8 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q5_K_S | llama.cpp/cuda | 198 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | H100 SXM | UD-Q4_K_M | llama.cpp/cuda | 188.1 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | H100 SXM | UD-Q5_K_M | llama.cpp/cuda | 185.3 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | H100 SXM | UD-Q4_K_XL | llama.cpp/cuda | 183.1 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | H100 SXM | UD-Q6_K | llama.cpp/cuda | 178.2 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q6_K | llama.cpp/cuda | 161.3 tok/s | 8192 | Structured table | source |
| Qwen3.6-35B-A3B | H100 PCIe | UD-Q4_K_M | llama.cpp/cuda | 152.6 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | H100 PCIe | UD-Q4_K_XL | llama.cpp/cuda | 152.1 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | H100 PCIe | UD-Q5_K_M | llama.cpp/cuda | 151.3 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | H100 PCIe | UD-Q6_K | llama.cpp/cuda | 146.1 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | RTX 5090 | UD-Q6_K_XL | llama.cpp/cuda | 139.6 tok/s | 8192 | Structured table | source |
More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.