Configs › Qwen3.5-35B-A3B tok/s: DGX Spark 131, Strix Halo 75.2
Qwen3.5-35B-A3B tok/s: DGX Spark 131, Strix Halo 75.2
Measured Qwen3.5-35B-A3B speeds: 131 tok/s on DGX Spark (NVFP4, atlas (speculative decode, mtp-2)). 4 configs, each linked to its source, snapshot 2026-10-15.
Snapshot 2026-10-15 · 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 Qwen3.5-35B-A3B on DGX Spark?
- 131 tok/s single-stream decode (NVFP4, atlas (speculative decode, mtp-2)), the only measured config on NVIDIA DGX Spark (GB10), measured by Atlas-Inf, source last checked 2026-09-23. source
- How fast is Qwen3.5-35B-A3B on Strix Halo?
- 75.2 tok/s single-stream decode (IQ4_XS, llama.cpp/vulkan), the fastest of 3 measured configs on AMD Strix Halo (Ryzen AI Max+ 395), measured by hogeheer, source last checked 2026-10-14. source
- How fast is Qwen3.5-35B-A3B on H100 SXM?
- 185.2 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA H100 SXM (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.5-35B-A3B on RTX 5090?
- 174.2 tok/s single-stream decode (UD-Q4_K_L, llama.cpp/cuda), the fastest of 18 measured configs on NVIDIA RTX 5090 (32GB), measured by paulplee, source last checked 2026-10-15. source
- How fast is Qwen3.5-35B-A3B on H100 PCIe?
- 151.3 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA H100 PCIe (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.5-35B-A3B on A100 40GB?
- 122.8 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA A100 40GB, measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.5-35B-A3B on A10?
- 97.9 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA A10 (24GB), measured by sypherin, source last checked 2026-10-14. source
Measured configs for Qwen3.5-35B-A3B on AI boxes
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.5-35B-A3B | DGX Spark | NVFP4 | atlas (speculative decode, mtp-2) | 131 tok/s | Extracted, source-linked | source | |
| Qwen3.5-35B-A3B | Strix Halo | IQ4_XS | llama.cpp/vulkan | 75.2 tok/s (prefill 1170.3) | 512 | Structured table | source |
| Qwen3.5-35B-A3B | Strix Halo | Q4_K_M | llama.cpp/vulkan | 64.9 tok/s (prefill 1080) | 512 | Structured table | source |
| Qwen3.5-35B-A3B | Strix Halo | Q4_K_M | llama.cpp/rocm | 54.7 tok/s (prefill 1047) | 512 | Structured table | source |
Qwen3.5-35B-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 |
|---|---|---|---|---|---|---|---|
| Qwen3.5-35B-A3B | H100 SXM | Q4_K_M | llama.cpp/cuda | 185.2 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-Q4_K_L | llama.cpp/cuda | 174.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-IQ3_XXS | llama.cpp/cuda | 169.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-IQ3_S | llama.cpp/cuda | 167.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-Q3_K_XL | llama.cpp/cuda | 167.1 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q3_K_M | llama.cpp/cuda | 166.9 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-IQ4_XS | llama.cpp/cuda | 166.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q4_K_M | llama.cpp/cuda | 166.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-IQ4_NL | llama.cpp/cuda | 166.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q4_K_S | llama.cpp/cuda | 166.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | MXFP4 | llama.cpp/cuda | 165.5 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-Q6_K_S | llama.cpp/cuda | 165.5 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-Q4_K_XL | llama.cpp/cuda | 164.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q3_K_S | llama.cpp/cuda | 163.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q6_K | llama.cpp/cuda | 163.5 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q5_K_S | llama.cpp/cuda | 163 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | Q5_K_M | llama.cpp/cuda | 161.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-Q5_K_XL | llama.cpp/cuda | 161.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | H100 PCIe | Q4_K_M | llama.cpp/cuda | 151.3 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.5-35B-A3B | RTX 5090 | UD-Q6_K_XL | llama.cpp/cuda | 127.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-35B-A3B | A100 40GB | Q4_K_M | llama.cpp/cuda | 122.8 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.5-35B-A3B | A10 | Q4_K_M | llama.cpp/cuda | 97.9 tok/s | 8192 | Verified (ran it) | 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.