Strata Runs DeepSeek-V4.1-Flash On A 64GB PC
Strata's pitch is that memory and storage, not VRAM alone, decide whether a 763-billion-parameter model can run at home, and the measurement puts that claim in numbers, with several hundred GB of disk as the practical catch.
Reporting from 1 source: ASCII.jp.
Niko Veit, developer of the open-source inference engine Strata, said on October 24 that the DeepSeek-V4.1-Flash model now runs on a PC with 64GB of main memory. The quantized release uses the GSQ-RCO compression method from DASLab at the Institute of Science and Technology Austria. Strata spreads model data across VRAM, main memory, and SSD. The 2-bit build measured 4 to 6 tokens per second on a GeForce RTX 5070 with 64GB of memory.
The provider of the quantized builds says the 2-bit version matched the original model's average score on three evaluations: math, science, and code generation. That claim matters more than the token rate, because a checkpoint that loses accuracy at 2 bits would make the compression interesting and not much else.
The three published models average 1.75, 2, and 2.25 bits of weight precision. Their download sizes, related files included, are about 239GB, 257GB, and 274GB. The 2-bit version needs roughly 265GB of free space. The measurement used Strata 0.2.5 with the 12GB of dedicated memory on a GeForce RTX 5070 and 64GB of main memory, and it generated 4 to 6 tokens per second.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google