Persistent memory and MCP tools make 27B enough for a local assistant on 24 GB

The 24 GB in the test rig is two separate 12 GB pools ^[26]. A model that fits in 12 GB runs on one card. A model that does not gets split, and the layers on the second card reach the first over the bus. Kernbach gives the cost of that split as roughly half the compute throughput from parallelization, plus the drop from x16 to x8/x8, which takes each card from 63 GB/s to 31.5 GB/s on PCIe 5.0 ^[10]. His machines are PCIe 4.0, where x8 is about 15.75 GB/s ^[17], a quarter of the PCIe 5.0 x16 number ^[23]. In one of the two configurations he tested, a 4070 sits in each PC and the split crosses the NVIDIA PAIR link between machines instead of the bus ^[16].

He expects the models to spend most of their time idle, not at 100 percent load ^[8]. The hardware choice rests on that assumption. He also says plainly what the build is for: the focus is the agent's ability to control and use integrated tools, and a local assistant is not a replacement for a cloud chatbot ^[21].

"A local model becomes genuinely useful not when it matches a cloud model in raw intelligence," Kernbach wrote, "but when it has persistent memory, access to tools, and the ability to interact with the user's real environment" ^[5].

The sufficiency claim covers 9B to 27B, under the 35B ceiling he set for himself ^[7]. The post reports power draw, temperatures and bus bandwidth; it does not report tokens per second, benchmark scores or task success rates ^[25]. His task list runs to documents in several formats with cross-document graphs, Google Calendar, Gmail and Drive, and Telegram both as a bot and in his own name. It also takes in external databases, smart home and sensor reads, and voice in his native language including transcription and translation ^[20]. In my view, picking the right tool call and filling its arguments correctly is an easier job for a model than unfamiliar reasoning, which is most of what this assistant does.

The card selection is a cost argument. Kernbach puts the 5080-to-5090 step at about 4x the price for less than 2x the performance and VRAM ^[11]. Divide one by the other and the price per unit more than doubles ^[24]. He points instead at the Ti and Super versions of the 5060, 5070 and 5080, with memory configurations up to 48 GB ^[12]. He recommends two cards from the same generation, because Windows drives both from a single driver ^[13].

Two cards at the measured draw is 220 to 240 W of GPU power ^[22]. His projections for full compute load run 800 to 900 W for two 5070 Ti cards with an i7-class CPU, close to 1 kW for two 5080s, and 1.4 to 1.5 kW for two 5090s. The same system with nothing to compute draws 50 to 70 W ^[14]. An assistant that idles most of the day costs that 50 to 70 W to keep running. In a vertical case the lower GPU warms the upper one, so he prefers a horizontal layout with good airflow ^[19].