RTX 4090 24GB for local AI in the UK
Ranks #19 of 23 for 7-9B models on value (£13 per token/s at £2,200 used).
Key facts
- Cheapest UK price
- £2,200 used
- Memory
- 24 GB
- Memory bandwidth
- 1008 GB/s
- Largest model it can run
- ~36B parameters at 4-bit (estimate: ~0.6 GB per billion parameters + ~2 GB overhead)
- Shop
- eBay UK (used)
- £2,200 used: median of 126 UK eBay sold listings, 90 days, checked date unknown (source)
Disclosure: eBay "Shop" links are affiliate links (marked rel="sponsored"): if you buy, we may earn a small commission at no extra cost to you. Rankings are never affected by commission.
Speed and value by model size
| Size class | Fits at 4-bit? | Speed | £ per tok/s | Value rank | Reference model |
|---|---|---|---|---|---|
| 7-9B models | Yes | 167.7 words/sec* from 22 tests other models: 145.8 (from 4 tests) | £13 | #19 of 23 | Llama 2 7B Q4_0 |
| 12-14B models | Yes | 97.7 words/sec* only 1 test ⚠ other models: 89.7 (from 4 tests) | £23 | low sample (n=1) ⚠, not ranked | Gemma 4 12B Q4_0 |
| 27-35B models | Yes | 47.1 words/sec* only 1 test ⚠ other models: 43.5 (from 3 tests) | £47 | low sample (n=1) ⚠, not ranked | Qwen 3.6 27B Q4_K_M |
| 70B+ models | No (needs ~42 GB+) | – | – | – | no reference model yet |
Speed is the median single-user decode speed running the like-for-like reference model for that class, in tokens per second (*a token is roughly ¾ of a word, so words/sec is a rough guide). Ranks use the cheapest UK price and need at least 3 test runs; ⚠ = low sample, shown but not ranked. "Fits" means the model fits fully in memory at ~4-bit. How we measure.
Other models measured
- llama3.1 Q4_K_M (8.0B, 7-9B class): 145.8 tok/s only 2 tests ⚠ · run
- qwen2.5-coder-14b Q4_K_M (14.0B, 12-14B class): 89.9 tok/s only 1 test ⚠ · run
- qwen2.5-coder Q4_K_M (14.8B, 12-14B class): 110.5 tok/s from 4 tests · run
- gpt-oss-20b Q4_K_M (20.0B): 242.7 tok/s only 1 test ⚠ · run
- lfm2.5-24b-a2b Q4_K_M (24.0B, 27-35B class): 231.9 tok/s only 1 test ⚠ · run
- Qwen 3.8 27B Q4_K_M (27.0B, 27-35B class): 43.1 tok/s only 1 test ⚠ · run
- qwen3.8-27b-mtp-on Q4_K_M (27.0B, 27-35B class): 84.5 tok/s only 1 test ⚠ · run
- gemma3 Q4_K_M (27.4B, 27-35B class): 92.7 tok/s from 3 tests · run
- glm-4.7-flash Q4_K_M (29.9B, 27-35B class): 164.9 tok/s only 1 test ⚠ · run
- qwen3-coder-30b-a3b Q4_K_M (30.0B, 27-35B class): 232.5 tok/s only 1 test ⚠ · run
- qwen3-coder Q4_K_M (30.5B, 27-35B class): 145.2 tok/s only 1 test ⚠ · run
- Qwen 3 32B Q4_K_M (32.0B, 27-35B class): 43.5 tok/s only 1 test ⚠ · run
- qwen3.6-35b-a3b Q4_K_M (35.0B, 27-35B class): 160.6 tok/s only 1 test ⚠ · run
- nvidia-nemotron-3.5-bolt-74b-a9b-24gb (74.0B, 70B+ class): 107.3 tok/s only 1 test ⚠ · run
- nvidia-nemotron-3.5-bolt-74b-a9b-24gb-ctx16k (74.0B, 70B+ class): 105.7 tok/s only 1 test ⚠ · run
- nvidia-nemotron-3.5-bolt-74b-a9b-24gb-ctx8k (74.0B, 70B+ class): 107.0 tok/s only 1 test ⚠ · run
Shown for information only: these don't affect rank. Mixture-of-experts models (names with "a3b", "moe" and similar) run much faster than their size suggests.
Similar price alternatives
- RTX 5080 16GB £1,276 used · 16 GB · 183.8 tok/s on 7-9B · #18 of 23 for 7-9B
- Mac mini M5 Pro (base config) £1,699 new · 24 GB · 63.9 tok/s on 7-9B · #21 of 23 for 7-9B
- AMD Strix Halo (Ryzen AI Max+ 395) 128GB £2,830 new · 128 GB · 50.6 tok/s on 7-9B · #23 of 23 for 7-9B
- Mac Studio M5 Max 40-core GPU (base config) £3,099 new · 48 GB · 111.6 tok/s on 7-9B · #22 of 23 for 7-9B
See the RTX 4090 24GB on the leaderboard → · All cards · Beginner's guide · Methodology
Weekly UK value moves, by email
Price drops, new benchmark results, best buys under £500 / £1,000 / £2,000.
Check your inbox to confirm: we use double opt-in, so you're only added after clicking the link in the confirmation email. Free; unsubscribe any time. Sent via Mailchimp (see their privacy notice).