About
A fit check for running open models at home.
RigFit compares open-weight models with consumer GPUs, Apple Silicon Macs, and a few mini PCs. Each model page shows an estimated memory need at Q4, Q8, and FP16. Each hardware page shows which of those models fit, the sticker memory and bandwidth, a list price, and a search link if you want to check a current price.
Who it is for
It is for people deciding whether a model will load on the machine they already have, and for people comparing machines before they buy one. That includes AI hobbyists running models locally, developers picking a quant for a laptop or a desktop, and hardware buyers who want memory and bandwidth next to a list price.
Why it exists
Memory calculators will tell you that a model needs a certain number of gigabytes. They usually stop there. They do not name the GPUs, Macs, and small desktops that actually have that memory, and they do not put a list price beside the spec. RigFit is that missing second half: a hardware pick, with the number the vendor published, and a way to go look up what it costs today.
How the numbers are calculated
Every fit figure is an estimate. None of them are a measured load or a benchmarked tokens per second.
When no GGUF file size is stored, estimated memory is the parameter count times the bytes per weight, plus 20% for a modest KV cache and runtime overhead. Q4 uses half a byte per weight, Q8 uses one byte, and FP16 uses two. Mixture-of-experts models are sized by the full weight set, not the active parameters. FP16 always uses this formula, because the catalog does not store an FP16 file size.
When a GGUF repo on Hugging Face names the catalog model as its base_model, the refresh script stores the Q4_K_M, Q5_K_M, Q6_K, and Q8_0 file sizes it actually finds. Q4 fit checks then use the Q4_K_M file, and Q8 fit checks use Q8_0. File size is bytes divided by 1e9 (decimal GB). A quant split into part files is summed. If that file is missing, the formula above is the fallback, and the page says so.
KV cache, when shown, is an FP16 estimate from public config.json fields, at 8,192 and 32,768 tokens. For GQA that is 2 × layers × KV heads × head dimension × tokens × 2 bytes. Hybrid models count only full-attention layers. Linear-attention layers are omitted, because they keep a fixed state this catalog does not size. For MLA (a compressed latent, used by some DeepSeek and GLM models) the estimate is layers × (kv_lora_rank + RoPE key dimension) × tokens × 2 bytes. If config.json also sets a sliding window, the number still assumes every counted layer stores the full context, so it can overstate those layers. Gated configs are left blank rather than filled from memory.
A GGUF fit adds the 8,192-token KV estimate to the file size. The 20% formula is not added on top of a file size.
Tokens per second, when shown, are the device bandwidth in GB/s divided by the active weight gigabytes read for one token. Mixture-of-experts models use the active-parameter share of the weight file. This is a bandwidth roofline, not a measured generation speed. Bandwidth marked “up to” is the vendor maximum, so the estimate is high for binned chips.
Discrete GPUs are treated as 95% usable, Macs as 70% of unified memory, and the Strix Halo mini PCs as 75%. A result marked Tight has less than 2 GB to spare. If Q4 weights exceed usable VRAM, the page states how many gigabytes would have to sit in system RAM. The catalog does not record system RAM, so that overflow is not called a fit. Unified-memory machines have one pool, so they are not described as offloading to a second one.
Read the methodology on the home page
Affiliate disclosure
RigFit participates in the Amazon Associates Program, an affiliate advertising program designed to provide a means for sites to earn advertising fees by advertising and linking to Amazon.com. “Check price on Amazon” opens an Amazon search. If you buy through that link, RigFit may earn a commission, at no extra cost to you. The link is not a live price, a specific offer, or a recommendation of a particular seller. Dollar figures on hardware pages are a launch MSRP or another cited list price. Price notes on each row say where that number came from.
Who runs it
RigFit is an independent site run by one person. There is no company, no staff page, and no outside editorial board. Hardware suggestions, corrections, and sponsorship inquiries go to the address on the contact page. A sponsored placement, if one is ever accepted, will be labeled as such.