Research
Research tied to real systems.
We investigate technologies, markets and operating problems when the answer is not obvious enough to justify immediate implementation.
01Fields of investigation
Where we look.
Research here exists to answer questions raised by engineering work. Every field on this list is one where we build, operate or actively investigate systems.
Computing Infrastructure Automation Industrial systems Energy Resource efficiency Logistics Security Economic systems
02Structure
Every entry follows the same structure.
A published entry is a working document, not an article. It states the question, the evidence, and what we currently believe. Entries are revised when the evidence changes. The goal is useful technical research, not commentary.
03Published entries
Evidence before opinion.
- Field
- Computing / Infrastructure
- Status
- Experimental result
- Hardware
- RTX 5090 / RTX 4090
- Largest test
- 671B parameters
- Result
- Functional inference beyond available VRAM
How far can large mixture-of-experts language models run on consumer GPU hardware?
Why it matters
Inference cost and data control push organizations toward local deployment, and most inference runtimes assume datacenter hardware: high-VRAM accelerators and uniform memory. If large mixture-of-experts models can be served well on consumer GPUs, local deployment becomes practical for organizations that cannot send data out, or cannot justify datacenter pricing.
Evidence & experiments
We built a local inference runtime from scratch in CUDA C++ and measured it on consumer hardware. The results below are measured, not projected.
- 86.8tok/sDense 8B model at Q8 on a single RTX 5090.
- 52.2tok/s30B mixture-of-experts model served entirely from a single consumer GPU.
- 23.3tok/s397B-parameter MoE across a dual consumer-GPU setup (RTX 5090 + RTX 4090), with the active expert set fully resident. Residency, not parameter count, sets throughput.
- 12.8tok/sSustained throughput at a 128K-token context.
- 671BparamsA 671B-parameter model runs on a single RTX 5090. Throughput is storage-limited at that size, but the run is functional.
- SSM+MoERuns a hybrid SSM+MoE checkpoint on consumer hardware.
Current conclusion
Consumer hardware can serve far larger models than commonly assumed, provided the runtime is designed around the memory hierarchy it actually has rather than assuming a datacenter card. Available VRAM is not a hard upper bound on parameter count.
Open questions
- Throughput ceilingsHow far the approach can be pushed before storage bandwidth dominates latency.
- Multi-tenant servingWhether the same strategies hold when several workloads share one GPU.
- Quantization quality boundsWhere aggressive quantization begins to cost measurable task accuracy.
Further entries will be published under the same identifier scheme, ALIA RESEARCH / 2026-NNN, when there is enough evidence to make the write-up useful.