Homelabber Runs 2.8T-Parameter Model on Optane Build

Adding two RTX 3090s and a RAM drive lifts Moonshot Kimi K3 generation speed on a home Optane system.

Updated on Nov. 21, 2026 in Artificial Intelligence

Homelabber Runs 2.8T-Parameter Model on Optane Build

A homelabber added two RTX 3090 GPUs to a second-generation Optane AI build running the 2.8-trillion-parameter Moonshot Kimi K3 model, reaching 3.11 tokens per second. Moving 45 of the model's 94 shards onto a RAM drive raised generation speed by 41%.

Why it matters

The results show that a very large model can run on consumer hardware by splitting its data between GPU VRAM, Optane memory and system RAM. The original build was designed to save money by using Optane instead of additional DRAM.

Two RTX 3090s provide 48 GB of VRAM, while the 865 GB model file is split so that 414 GB (45 of 94 shards) sits on a RAM drive. The routed-expert tensors remain on the CPU side in Optane plus the RAM drive, and the GPUs handle attention, the dense feed-forward layer, shared experts and the KV cache.

The details

The build runs the 2.8-trillion-parameter Moonshot Kimi K3 model with GPU and CPU-side memory divided by tensor type. The homelabber reports that more optimizations are possible, including customized quantization, and that per-tensor or per-layer quantization might make the setup faster still.

Timeline

  1. November 21, 2026: Article published on the homelabber's GPU-upgraded build.

The Tech Race

The build relies on Optane, a memory technology chosen to avoid the expense of extra DRAM, and this result shows that a cost-saving design can still run a trillion-parameter model at usable speeds. It positions home AI hardware as a practical alternative to expensive data-center memory for running very large open-weight models.

The results suggest that hobbyists with several GPUs and a large amount of RAM could run very large models locally, though the setup requires considerable hardware and tuning. Speed gains came from moving selected model shards to a RAM drive and may depend on the specific machine.

The takeaway

Hardware choices such as Optane memory, GPU VRAM and RAM placement can significantly change how fast a very large model runs at home. Readers interested in local AI can test similar tradeoffs on their own systems before buying new equipment.

Further reading

Explore more on Artificial Intelligence to see how large models are being run on local hardware.

Source note: This article includes information reported by Tom's Hardware.