RigValue UK

🏆 Leaderboard: £ per token/s

Hides options whose UK price is above this amount, in both the picks and the table. Leave blank for no cap.

Picks the biggest model size class your card can run with measured data, and highlights your card.

How to read this: Speed is how fast the AI types its answer, in tokens per second (a token is roughly ¾ of a word, so *words/sec is a rough guide). "From 111 tests" means 111 real people's test runs back up that number, so more tests means you can trust it more. £ per speed is the price divided by the speed, and lower is better value.

Technical detail: Lower £ per tok/s = better value. Speed = median single-user decode tok/s across all measured dense models of that size class at ~4–6-bit (MoE models such as 35B-A3B are excluded because they are far faster than their size suggests). Only runs where the whole model fits in card/unified memory count (partial CPU offload excluded). n = number of runs; ⚠ = a single run, not ranked unless you tick the box. Sources: llm-speed.com "chat-short" runs and llama-bench tg posts in llama.cpp GitHub discussions. Blank = no verified data yet, never estimated.

Every option has its own page with prices, speeds by model size and similar-price alternatives: tap its name, or browse all cards. Shopping to a budget? See the best picks under £300, under £500, under £1,000 and under £2,000.

💷 What can I run for £X?

"Largest Q4 model" is a rough rule of thumb (~0.6 GB per billion parameters plus ~2 GB overhead), labelled as an estimate.

🔒 Why run it yourself

Every question you type into a cloud AI is sent to someone else's servers, where it can be logged, reviewed, used for training or handed over. A model running on your own hardware doesn't send anything anywhere.

• Your data stays with you. Your prompts, documents and code never leave your house, and it still works with the internet off.
• Nobody can change it on you. No surprise price rises, model swaps, rate limits or new terms. The version you download stays the version you have.
• No bill per question. You pay once for the hardware, and after that each answer costs only electricity.

That's why we rank hardware by value: owning your AI should be affordable, not a luxury.

🧭 New to running AI locally?

The easiest way to start, step by step: a free phone app with zero setup, or one free computer app and a small model, chatting in about 15 minutes. Plus what works easily on what you already have.

Read the beginner's guide →