Contributors › julianmb's local-LLM benchmark configs

julianmb's local-LLM benchmark configs

Every tracked benchmark config measured by julianmb: model, quant, backend, tokens per second, source-linked and trust-tiered.

Snapshot 2026-10-30 · 412 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.

18 configs · 6 models · 211 stars · 105.6 tok/s best single-stream · julianmb on github.

Configs measured by julianmb

Model	Hardware	Quant	Backend	Decode	Context	Trust	Source
Ornith-1.5-35B-A3B	Strix Halo	FP4	speculative decode, mtp	105.6 tok/s		Owner-submitted	source
NVIDIA Nemotron 3.5 Lightning 30B	Strix Halo	FP4	speculative decode, mtp	95.2 tok/s		Owner-submitted	source
Nex N2.5 Mini 35B-A3B	Strix Halo	FP4		76.9 tok/s		Owner-submitted	source
Ornith-1.5-35B-A3B	Strix Halo	FP4		76.9 tok/s		Owner-submitted	source
Nex N2.5 Mini 35B-A3B	Strix Halo	Q4_K_M		72.8 tok/s		Owner-submitted	source
Ornith-1.5-35B-A3B	Strix Halo	Q4_K_M		71.6 tok/s		Owner-submitted	source
Qwen3.8-27B	Strix Halo	FP4	speculative decode, mtp-4	33.8 tok/s		Owner-submitted	source
DeepSeek V4 Flash 284B	Strix Halo	IQ2_XXS	speculative decode, mtp	32 tok/s		Owner-submitted	source
Qwen3.8-27B	Strix Halo	FP4	speculative decode, mtp	26.9 tok/s	32768	Owner-submitted	source
Qwen3.8-27B-DFlash2	Strix Halo	Q4_K_M	vulkan (speculative decode, dflash2)	21.2 tok/s (prefill 211)	32768	Owner-submitted	source
Qwen3.8-27B	Strix Halo	FP8	llama.cpp (speculative decode, mtp)	19 tok/s		Owner-submitted	source
Qwen3.8-27B	Strix Halo	FP4	llama.cpp/rocm	14 tok/s		Extracted, source-linked	source
Qwen3.8-27B	Strix Halo	FP4	vulkan	13.9 tok/s (prefill 266.6)	16384	Owner-submitted	source
Qwen3.8-27B	Strix Halo	Q4_K_M		12.4 tok/s		Owner-submitted	source
Qwen3.8-27B	Strix Halo	Q4_K_M	llama.cpp	12.3 tok/s	32768	Owner-submitted	source
Qwen3.8-27B	Strix Halo	Q4_K_M	llama.cpp/rocm	12.3 tok/s		Extracted, source-linked	source
Qwen3.8-27B	Strix Halo	FP8	llama.cpp/rocm	7.7 tok/s		Extracted, source-linked	source
Qwen3.8-27B	Strix Halo	FP16	llama.cpp	5 tok/s		Owner-submitted	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.