Qwen3.8 27B Runs From a 12 GB Laptop by Pooling Memory Across 4 Devices

A 12 GB Windows laptop has successfully served as the primary system for running Qwen3.8 27B by distributing the model across 4 completely different devices on the same local network. The experiment combines the laptop with an RTX 3060 mini PC, a Mac mini and an Android smartphone, demonstrating how unused RAM and VRAM across existing hardware can be turned into a distributed local AI system instead of requiring a single machine with enough memory to hold the entire model.

The experiment was published by Reddit user Medicine_Blogscanner in the original demonstration. The 12 GB Windows laptop remained the primary node and allocated around 3.4 GB to the workload. The RTX 3060 mini PC contributed approximately 20 GB of combined system memory and VRAM, while the Mac mini provided 3.7 GB and the Android smartphone contributed another 1 GB. Together, the devices were able to load and run the 27B model despite the primary laptop being unable to accommodate it alone.

The software behind the setup is RAMDeck, which distributes large language model inference across heterogeneous hardware. Its RAMDeck Core repository explains that the engine shards LLM inference across Mac, PC and Linux systems and communicates through the standard llama.cpp RPC protocol. One technical distinction is worth noting: RAMDeck is described by its developers as source available rather than OSI approved open source because its Apache 2.0 license includes a Commons Clause restriction for commercial resale.

The result is technically impressive, but performance makes the limitations of this approach immediately clear. Qwen3.8 27B generated approximately 1.92 tokens per second during the test, with the reported benchmark showing around 25 ms latency. The developer previously tested the same distributed concept with smaller 7B and 13B models, with performance declining as model size increased. The next test is expected to move the primary role from the older laptop to the more capable RTX 3060 mini PC to examine how much the host system affects loading and inference speed.

The memory challenge becomes easier to understand when looking at the model itself. The official Qwen3.8 27B model contains 27 billion parameters, while its full model repository occupies roughly 55.6 GB. Qwen lists a native context length of 262,144 tokens with support for extension to 1,000,000 tokens. Reduced precision quantization can dramatically lower the amount of memory required to run it, but larger versions can still exceed the memory available in ordinary consumer laptops and graphics cards.

RAMDeck does not turn the memory inside several machines into one physically unified memory bank. Instead, different portions of the workload remain distributed across individual systems and must communicate over the network. That creates a major performance disadvantage compared with keeping model weights and processing close to a GPU through high bandwidth local memory. The experiment therefore solves the capacity problem far more effectively than the bandwidth problem.

That distinction is becoming increasingly relevant as local AI hardware moves toward much larger memory pools. Systems such as the NVIDIA RTX Spark platform are moving toward as much as 128 GB of unified memory specifically to keep large AI workloads close to the available compute resources. NVIDIA is also developing PAIR around connecting multiple local AI systems, showing that distributed consumer AI compute is becoming a serious area of development rather than simply a home laboratory experiment.

Running Qwen3.8 27B at 1.92 tokens per second is not going to replace a dedicated AI workstation, but raw speed is not what makes this experiment interesting. The important part is proving that memory capacity can be assembled from hardware users already own. An old laptop, a gaming mini PC, a Mac and even a smartphone can contribute toward the same inference workload.

The obvious bottleneck is communication. AI workloads benefit enormously from memory bandwidth, which is why unified memory systems and large VRAM GPUs remain much faster solutions. Still, RAMDeck demonstrates another possible direction for local AI where existing machines become nodes rather than obsolete hardware. If software can reduce the networking and orchestration overhead, distributed home AI systems could become considerably more practical.

Would you combine several older PCs and devices to run larger AI models locally, or would you rather invest in 1 high memory AI system?