Four local LLMs up to 748B parameters benchmarked on a single RTX 5090

A tester ran Qwen3.8-Flash-Next (125B), Ornith-1.5-122B, GLM-5.3-Flash (321B) and DeepSeek-V4.1-Flash (748B) on one RTX 5090 with 128 GB of RAM and a Ryzen 9 9950X3D. Qwen on Strata hit 80.2 tok/s and 4,075 tok/s prefill on a ~31K prompt, while DeepSeek on Project Maya reached 21.8 tok/s and 622 tok/s. All four found the needle at contexts up to ~125K tokens.

• Qwen3.8-Flash-Next on Strata: 80.2 tok/s, 4,075 tok/s prefill at ~31K tokens
• DeepSeek-V4.1-Flash on Maya: 21.8 tok/s, 622 tok/s prefill, 3.07 s to first token
• GLM-5.3-Flash scored 105/108 on the task suite, Ornith scored 89/108
• All four MoE models fit in 32 GB of VRAM with experts offloaded to RAM