FAQ

What is the difference between Q4 and BF16?

BF16 (or FP16) stores each weight in about 2 bytes — close to the original training precision, and the largest download. Q4 (for example Q4_K_M) stores weights in about 4 bits, so the file is roughly a quarter the size and usually still good quality for chat. Q8 sits in between. Smaller quants (Q3, Q2, IQ2…) fit tinier cards but can lose more quality. We always show the real Hugging Face file size, not a parameter-count guess.

What is the KV cache?

Beyond the weights, the engine keeps a key/value cache for every token in the conversation (and the prompt). That cache grows with context length. A long context can add as much memory as the weights themselves. Each model page shows bytes-per-token from config.json when we can source it. If we can't, we say so and the fit table uses weights only.

Why do rental prices change?

Cloud GPU prices move by the hour. We copy the public list price on the page's "checked on" date. Always confirm on the provider's site before you rent. Affiliate or referral links never change the number we publish.

What does "checked on" mean?

It's the date we last verified that page's figures against the cited sources (Hugging Face file list, config, vendor specs, or pricing page). When we fix a wrong number, we bump the date.

Do you sell rankings or "Fits" verdicts?

No. See Disclosure. Money never changes a number.

How do I report a mistake?

Email vramneeded.partners [at] gmail.com — see About.