Start here
Guides
Two questions decide a local-AI build: does the model fit in memory, and does the machine around it keep up. These guides answer both with numbers from the catalogue we sell — no spec-sheet theatre, and no benchmark we have not run.
Buying guide · 4 chapters
Find hardware for my AI workload
Updated 01 Nov 2026
How much memory you actually need
The table that decides everything else: what 8, 24, 48 and 80 GB of VRAM actually run at 4-bit, and where the headroom goes.Choosing between accelerator classes
Memory bandwidth sets tokens per second, capacity decides whether the model fits at all — so pick the class before the model.Power, cooling and system balance
PSU headroom, airflow and the parts that decide whether a build holds full clocks under sustained load.Shortlisting from the catalogue
Turning a memory target into two or three listings you can compare side by side — and knowing when to stop optimising.
From the blog
Worked examples
Where the guides cover the rules, these posts run the arithmetic on a specific model — with prices, tokens per second and the point where the API wins after all.
12 Nov 2026
Qwen 4 27B: what the release date tells you about your next local LLM
The 27B tier is where local LLM work actually happens. Here is what Qwen 4 27B changes, what it probably won't, and how to size your rig before the weights drop.
Read10 Nov 2026
How much VRAM do you need? A 70B model takes 40 GB
A 70B model needs 35–40 GB before context. Your laptop has 16. Here's the memory arithmetic behind every model size, and how to pick one that actually fits.
Read07 Nov 2026
12 GB Is Enough for DeepSeek V4.1 Flash — With Conditions
The 12 GB claim is testable, not a guarantee. Here is what decides whether DeepSeek V4.1 Flash fits your card — and a five-step procedure to measure it yourself.
Read
Have a model in mind?
Tell us the model, the context length and the number of users and we will point at the two or three listings that fit — or at the honest answer that the API is cheaper for your volume.