Will it fit?
A model fits when everything it needs is smaller than the memory the GPU can use:
- Weights: the real size of the quantised file on Hugging Face (a Q4_K_M, a Q8_0, an MLX 4-bit build). Where no file size is published, the parameter count times the quant's bits per weight.
- KV cache: what the model remembers about the conversation, which grows with context length. It is worked out from the model's own configuration: layers, attention heads and head size, with sliding-window, hybrid linear and latent (MLA) attention each counted the way they actually store it. A q8 cache halves it, which is why some rows show a second, longer maximum context.
- Runtime: buffers and the driver's own share, from a per-platform profile.
Usable memory is the card's VRAM less what the display and driver keep. On Apple silicon it is the share macOS lets the GPU wire by default, about two thirds of memory up to 36 GB and three quarters above; a raised limit is shown separately as tuned. With two cards, layers are split between them.
Three answers come out: fits (entirely in GPU memory), offload (part of it spills into system RAM, which is much slower) and won't run.
How fast?
Generating each token means reading the active weights and the KV cache from memory once, so decode speed is set by memory bandwidth: efficiency × bandwidth ÷ bytes read per token. A mixture-of-experts model only reads its active experts, which is why a 35B-A3B can be quicker than a dense 14B. When layers are offloaded, the system RAM's bandwidth (channels × transfer rate) sets the pace for those layers.
Prompt processing (prefill) is set by compute instead: the GPU's FP16 throughput against two operations per active parameter per token, plus an attention term that grows with context. It decides how long you wait before the first word of an answer, which matters most for coding agents and long documents.
The share of peak bandwidth and compute each platform achieves with llama.cpp:
| Platform | Decode efficiency | Prefill efficiency | Mixture of experts | Reserved per GPU |
|---|---|---|---|---|
| NVIDIA (CUDA) | 62% | 45% | 60% | 0.6 GB |
| AMD (ROCm) | 55% | 35% | 60% | 0.8 GB |
| Vulkan | 60% | 30% | 60% | 0.8 GB |
| Apple (Metal) | 65% | 40% | 60% | – |
| Intel (SYCL) | 50% | 30% | 55% | 0.8 GB |
These start from published llama.cpp results and are refitted as reference measurements are added; each refit gets a new calibration version. Ollama and LM Studio use llama.cpp underneath and land close to it. vLLM and MLX are usually a little faster on the hardware they support.
Good for
A machine is called good for a use when the most precise quant that fits reaches these speeds at that use's context:
- chat: at least 15 tokens a second, at 8k context.
- coding: at least 20 tokens a second, at 32k context.
- agents: at least 30 tokens a second and 500 tokens a second of prompt processing, at 32k context. Only models trained for tool use count.
- long documents: at least 15 tokens a second and 300 tokens a second of prompt processing, at 128k context.
What the estimates cannot tell you
- Quality. A Q4 build of a model is a little worse than its Q8; we show which quant each speed is for, but not how much it matters for your work.
- Thermals, power limits and slower memory on laptop and small-form-factor parts.
- Batching. The speeds are for one conversation at a time, which is how most people run models at home.
- New architectures before llama.cpp and the other runtimes support them properly.
If your own numbers are very different, we want to know: owners' measured results are coming, and they will appear beside the estimates.
Prices
Prices are what retailers in the UK charge, in GBP and incl. VAT, collected from retailer APIs and affiliate product feeds several times a day. We never convert currencies silently. Each page says when prices were last checked, and a price we could not confirm recently is not shown.
- A price that moves more than 40% from its weekly median is held until a second check confirms it.
- A new card listed far below its launch price is treated as a probable mismatch or scam and hidden until someone checks it.
- Some retailers (Amazon, until its product API is available to us) are shown as a link without a price.
- Used prices are eBay asking prices, the middle of what sellers want today, not what cards actually sold for.
Retailer links are affiliate links: we may earn a commission at no extra cost to you. It never changes the order of anything; tables sort by price, speed or popularity, as labelled. See about readygpu.
The catalogue
Models, quantised builds and hardware are added every day. Architecture facts come straight from each model's configuration file; specifications come from manufacturers' pages and are cited on each product. An AI assistant helps find new releases and read spec sheets, but every change is checked by code, and new models and new hardware are reviewed by a person before they appear. Every change is listed on What's new.