Methodology
Under the hood
Speed
Speed: median decode tokens/sec of runs labelled measured on the chat-short workload (single user), grouped by hardware, model and quantisation. Each figure links to its source run.
Price
Price: used = median of UK eBay sold listings over 90 days (PC Atlas). New = PC Atlas "New from" column, or the Apple UK Store "From" price for the base configuration. Each price shows its source and date.
Size classes
Size classes: 7–9B (6–10.9B), 12–14B (11–15.9B), 27–35B (24–36.9B), 70B+ (65B+), by total parameters. We use dense models only (MoE excluded) at Q4–Q6 quantisation; F16/Q8 runs are left out. Each class is ranked on one reference model and quantisation (see below); other models are shown for context only. Options under the rough memory needed at 4-bit (6 / 10 / 20 / 42 GB) are excluded from that class, and the page says which ones. Figures from fewer than 3 tests are flagged ⚠ and not ranked by default.
Fits-in-memory rule
Only runs where the model fits fully in card or unified memory count: llama-bench posts need full GPU offload (ngl ≥ 99), and estimated weights (parameters × bits ÷ 8) plus ~1.5 GB must fit in memory. Runs with unknown memory are left out.
Card matching
Card matching: GitHub posts are matched by card name, and only to vendor-appropriate threads (e.g. no AMD/Apple hits from the CUDA thread). Laptop/mobile GPUs and multi-GPU posts are skipped. Variants with several memory sizes are separated: the 4060 Ti/5060 Ti need an explicit 16GB mention, and 3060 posts that say 8GB are dropped.
Reference model
Like-for-like ranking: different models and compression levels run at very different speeds, so mixing them would be unfair. In each size class we pick one reference model and quant: the one measured on the most options (currently Llama 2 7B at Q4_0 for 7-9B, and Qwen 3.8 27B at Q4_K_M for 27-35B). Rank uses only runs of that reference. Runs of other models/quants are still shown in grey as extra information but never change the rank. The reference for the selected class is named above the table.
3-test minimum
Minimum sample: an option needs at least 3 reference runs to get a rank. Options with fewer are marked ⚠ and only ranked if you tick "Include low-sample speeds".
Physics check
Physics check: generating each token means reading the whole model from memory, so decode speed can't exceed memory bandwidth (GB/s) ÷ model weight size (GB). We reject any run faster than that limit plus 10% headroom (usually a mislabelled card, multi-GPU setup or speculative decoding). Bandwidth uses published spec figures (Apple's stated bandwidth for Macs). Mixture-of-experts models are exempt because only part of the model is read per token. Rejected runs are listed in
data.json under physics_rejects.Mac memory
Mac memory: Mac prices are Apple UK's "From" price for the exact chip config, and memory is that config's base unified memory per Apple UK specs (Mac mini M5 Pro 24GB; Mac Studio M5 Max 40-core GPU 48GB, the minimum Apple allows for that chip; Mac Studio M5 Ultra 96GB).
£ per tok/s
£ per tok/s = price ÷ decode tok/s. Lower is better. GPU prices are for the card only, not a whole system.
Gaps
Gaps: if we have no verified price or no measured run, the cell stays blank. We never estimate or fill in speeds.
Refresh schedule
Refreshed daily by
scripts/refresh.py.Credits & licences
Sources
- Benchmark data: llm-speed.com open dataset, licensed CC BY 4.0. We filtered and aggregated it (medians by hardware/model/quant). These are submitted measurements and not independently verified.
- Benchmark data: llama-bench results posted in llama.cpp GitHub discussions #15013 (CUDA), #15021 (ROCm), #10879 (Vulkan) and #4167 (Apple Silicon). These are user posts (GitHub ToS); each figure links to its comment, credited to its author. Only tg (decode) rows are used. Hardware is matched from the post text, and multi-GPU posts are skipped.
- Model names are normalised across sources (file suffixes, quant labels, -instruct/-it) so the same model and quant compare directly.
- Planned: localai.computer dataset (CC BY 4.0 compilation). Not included yet because automated download is blocked by a browser check.
- UK used/new GPU prices: PC Atlas. Apple prices: Apple UK Store. Strix Halo price: GMKtec UK list price (GMKtec EVO-X2, 128GB).
Disclosure
Affiliates
eBay "Shop" links are affiliate links (eBay Partner Network). If you buy through them, RigValue may earn a commission at no extra cost to you. Rankings are never affected by commission: they are calculated only from price and measured speed.