ViableView internal session, 9 October 2026
Self Hosting LM Models: Concepts and Practices
Running language models on your own hardware, from how inference works to choosing a stack.
- Organizer
- ViableView
- Attendees
- 8
- Type
- Talk
- LLM
- self-hosting
- inference
- quantization

Abstract
An internal ViableView session on running language models on your own hardware. It covers how inference works, why ASICs win over CPUs and GPUs for this work, the provider landscape of VPS, hyperscalers, and neoclouds, how accelerators compare, the software stack from Ollama and MLX to vLLM, and quantization.
Outline
- How inference works: tensors, matrix multiplication, prefill, and decode
- CPUs, GPUs, and ASICs in deep learning
- The provider landscape: VPS, hyperscalers, and neoclouds
- How accelerators compare on memory, bandwidth, and price
- The compute stack: Ollama, MLX, and vLLM
- Quantization and inference