Running models on your own hardware
Open weights are free. The machine to run them is not. This section answers the five questions that decide whether self-hosting is worth it for you, with the numbers in tables rather than spread across a hundred forum replies.
- What configuration actually holds the model, card by card
- How much memory each model needs at each quantisation
- How fast it runs, and why capacity and speed are different problems
- How good the result is, expressed as the closed model it matches
- What it costs today, with the date the price was checked
Coding
Which GPU runs a coding model, what the ceiling is against Claude and GPT, why open weights often still need a server, and the cheapest hardware that actually works.
live8-12 GB cards
The playbook for production work on an ordinary GPU: which models fit, six deploy steps, and the honest pros and cons against a subscription.
liveImage
Local image generation: model sizes, VRAM per resolution, throughput, and what a workstation costs against a subscription.
plannedHardware and video
MTT S5000, Mac Studio Ultra 512GB and RTX Pro 6000: prices with source labels, and a VRAM fit matrix for LTX-2.5, HunyuanVideo 1.5 and Wan.
liveMusic and speech
Local audio generation and text to speech: model choices, memory, and real-time factors.
planned