Method
How we estimate fit and speed
These are back-of-envelope estimates, not benchmarks. They're useful for ruling machines in or out, not for exact numbers.
Prices and stock
A script checks each retailer page and records the price and stock text exactly as shown, with the date. If we couldn't see a price, we say “unknown” instead of guessing. GPU prices are for the card only. Prices exclude tax and shipping. Always confirm on the retailer's page before buying.
Will the model fit?
memory needed (GB) = weights + 1.5 GB + 5% of weightsweights = total parameters (billions) × bits per weight ÷ 8
Q4_K_M: 4.85 bits per weight. ~4.85 bits/weight average for llama.cpp Q4_K_MQ8_0: 8.5 bits per weight. ~8.5 bits/weight for llama.cpp Q8_0- The 1.5 GB + 5% term roughly covers the runtime plus the KV cache at about 8K tokens of context for one user. Long contexts (32K–1M) can need much more; models with sliding-window or compressed attention need less.
Usable memory is installed memory times a per-machine fraction. Discrete GPUs get about 93%, since the driver and display take the rest. DGX Spark/GB10 boxes get about 85%, because memory is shared with the OS. Macs and Strix Halo get about 75%: macOS and the AMD driver cap GPU-usable memory by default. You can raise that cap with tuning, so treat “Won't fit” near the line as “maybe”.
A model gets Fits under 90% of usable memory, Tight fit between 90% and 100%, and Won't fit above that. We don't count offloading to system RAM, which works but is much slower.
How fast will it run?
tokens/sec ≈ 0.6 × memory bandwidth (GB/s) ÷ (active parameters (B) × bits per weight ÷ 8)
Generating each token means reading every active weight from memory once, so bandwidth sets the ceiling. Real systems reach roughly 50–70% of peak; we use 60%. For mixture-of-experts (MoE) models, only the active parameters count. That's why a 120B MoE can be faster than a 70B dense model. This is decode speed for one user. Prompt processing (prefill) depends on compute, not bandwidth, and that's where NVIDIA GPUs and DGX Spark pull ahead of Macs and Strix Halo.
Stock labels
In stock the retailer shows it orderable now. Pre-order pre-order or a stated delay. Out of stock the retailer says out of stock. Stock unknown we couldn't read stock. Hover a label to see the retailer's exact wording.
Where the numbers come from
Memory bandwidth comes from manufacturer spec pages. Model parameter counts come from the model makers and public catalogs. Product images come from manufacturer and retailer product pages and press kits, and each source is recorded in our data file. Corrections are welcome: reply to any newsletter or email hello@bestlocalairig.com.
Contact, sponsorship & listings
Questions, corrections or press: hello@bestlocalairig.com.
Sponsor / list your hardware: if you make or sell local-AI hardware and want it listed or reviewed, or want to sponsor the newsletter, email hello@bestlocalairig.com. Sponsored placements are always labeled and never change the fit or speed numbers.
Affiliate disclosure
Some “Check price” links may become affiliate links. We'd earn a small commission if you buy, at no extra cost to you. That never changes the rankings, which come only from price, memory and bandwidth.