Contributors › sypherin's local-LLM benchmark configs
sypherin's local-LLM benchmark configs
Every tracked benchmark config measured by sypherin: model, quant, backend, tokens per second, source-linked and trust-tiered.
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.
16 configs · 10 models · 85 stars · 81 tok/s best single-stream · sypherin on github.
Configs measured by sypherin
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| llm-jp-4.1-32b-a3b-thinking | Strix Halo | Q4_K_M | llama.cpp/vulkan (depth-0) | 81 tok/s (prefill 1980.2) | Verified (ran it) | source | |
| Gemma-4-26B-A4B | Strix Halo | UD-Q4_K_XL | llama.cpp | 78 tok/s | Owner-submitted | source | |
| llm-jp-4.1-32b-a3b-thinking | Strix Halo | Q4_K_M | llama.cpp/rocm (depth-0) | 73.6 tok/s (prefill 1416.6) | Verified (ran it) | source | |
| llm-jp-4.1-32b-a3b-thinking | Strix Halo | Q4_K_M | llama.cpp/vulkan (depth-4096) | 72.6 tok/s (prefill 943.6) | 4096 | Verified (ran it) | source |
| Qwen3.6-35B-A3B-MTP | Strix Halo | UD-Q4_K_XL | vulkan (speculative decode, mtp-3) | 66 tok/s | 262144 | Verified (ran it) | source |
| llm-jp-4.1-32b-a3b-thinking | Strix Halo | Q4_K_M | llama.cpp/rocm (depth-4096) | 64.8 tok/s (prefill 1021.1) | 4096 | Verified (ran it) | source |
| Qwen3.8-Flash-Next | Strix Halo | UD-IQ4_XS | halogen/rocm (speculative decode, mtp) | 63.7 tok/s (prefill 1545.4) | 131072 | Verified (ran it) | source |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q8_K_XL | llama.cpp/vulkan | 41 tok/s | Verified (ran it) | source | |
| Qwen3.6-35B-A3B | Strix Halo | UD-Q8_K_XL | llama.cpp/rocm | 39 tok/s | Verified (ran it) | source | |
| Qwen3.8-27B | Strix Halo | FP4 | llama.cpp/rocm (speculative decode, mtp-4) | 36.8 tok/s (prefill 284.5) | 131072 | Verified (ran it) | source |
| Muse-Glimmer-30B | Strix Halo | FP4 | rocm (speculative decode, dflash) | 26.5 tok/s | Owner-submitted | source | |
| Qwen3.8-27B | Strix Halo | UD-Q4_K_XL | llama.cpp/vulkan (speculative decode, mtp) | 24.6 tok/s | 131072 | Verified (ran it) | source |
| Qwen3.8-27B VL | Strix Halo | UD-Q4_K_XL | llama.cpp (speculative decode, mtp) | 22 tok/s | Owner-submitted | source | |
| Qwen3.6-27B | Strix Halo | UD-Q4_K_XL | llama.cpp (speculative decode, mtp-5) | 21 tok/s | 262144 | Owner-submitted | source |
| Qwen3.8-27B | Strix Halo | Q8_0 | vulkan (speculative decode, mtp) | 17 tok/s | 131072 | Owner-submitted | source |
| DeepSeek-V4-Flash-0731 | Strix Halo | UD-IQ2_XXS | vulkan | 12 tok/s (prefill 130) | 131072 | 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.