eleticle

Core No. 07Prices 15 Nov 2026Code & data →

What hardware does a mineral-processing site need to self-host LLMs and expert models?

Calculator · six quantised models · Australian prices incl. GST · speeds fitted to 784 first-hand runs

A calculator for a site's own AI machine. Pick a language model, a machine, how long each conversation is and how many answers it writes at once; it estimates the speed and the hardware cost, beside real runs on similar machines.

What we found

  • One card can run a big modelAn A$43,568 machine with one 96 GB card, plus a 16 GB tools card, runs GLM-5.3-Flash, a 320-billion-parameter model, at an estimated 56 tokens per second (35–100) for one 128K-token conversation: its experts live in RAM as well as on the card.
  • RAM holds the big modelsThe large models spread their experts over the card and RAM, the busiest on the card.
  • The software stack mattersOn one computer with the same model file, a stack built for these models, with its default drafting, wrote three to five times faster than a general one without.
  • Two jobs, two cardsMain cards for the language model; a 16 GB card for vision and embeddings.
  1. Method
  2. 01Your machine
  3. 02Three machines
  4. 03The stack
  5. 04Limits

Method

How the numbers are made

  • Measured784 first-hand runsOne stream each: 555 writing and 363 reading speeds, only 33 of them writing at 128K or more. From GitHub, X and the LocalMaxxing leaderboard, each with its hardware and settings.
  • EstimatedWhere every byte livesEach model’s file split into experts, dense weights and lookup tables, placed in VRAM, RAM or SSD, then read at each part’s speed.
  • FittedA few constants per stackFitted to the runs. Runs on other kinds of hardware, with other drafting settings or with too little RAM are kept but not used.
  • PricedAustralian retail, incl. GSTCards, hosts, RAM, power and network as listed on 15 November 2026. Prices change.

Prices, file sizes, stack versions, every run with its conditions, the fitting script and its error on every run are in the article’s public repository, where the stacks are also named. Recheck prices before buying.

01 · Your machine

Pick a model and a machine

Change the model, the machine, the context or the number of people writing at once. The estimate and the runs measured on similar machines change with it.

The Medium machine (PRO 6000 Max-Q 96 GB · used EPYC, 256 GB ECC) running GLM-5.3-Flash, one stream at 128K context. The calculator needs JavaScript.

Hardware cost, incl. GST

A$43,568

1 × PRO 6000 Max-Q A$24,999 · tools card A$2,299 · host A$14,700 · power and network A$1,570

Listed prices of 15 Nov 2026; the card from e365.com.au. On 22 Nov 2026 it was A$27,999 on backorder: A$46,568 in all.

  • Writing each stream35–100tokens/s · estimate
  • Reading a cold prompt844–4,056tokens/s · a full 128K prompt in 32–154 s
  • Streams at your speed4each about 20 tok/s or more at 128K, mid estimate; 2 at the low end
  • Longest context1024Kthe model’s own limit
Each stream’s writing speed as more run at once
Each stream writes about 56 tokens per second alone, falling as more run at once; at the middle of the range 4 keep 20 tokens per second each, at its low end 2.
Where the model lives
VRAM holds the dense weights, the context and 93 GB of experts; RAM holds the rest of the experts.

In GB. Cards and RAM are sold in GiB: 24 GiB is 25.8 GB. Each card keeps 1.6 GB for the runtime.

Measured on similar machines

    02 · Three machines

    A small, a medium and a large machine

    Each with its 16 GB tools card, each model in the largest quantised version that fits and keeps 20 tokens per second, one stream at 128K context. Load one into the calculator to change it.

    Estimated writing speed of each model on the Small, Medium and Large machines at 128K context.

    Estimated tokens per second of one answer. Each bar is the middle of a range; load a machine to see the range and the runs behind it.

    03 · The stack

    For the big models, the stack matters as much as the card

    Measured, not estimated. Stacks built for mixture-of-experts models keep the busiest experts on the card and the rest in RAM. A general GGUF runner splits the model by layers instead.

    Measured: on one computer, an expert-cache stack with its default drafting wrote three to five times faster than a general GGUF runner without drafting; on two different machines, three times.

    The first two are one computer, measured by the expert-cache stack’s author: it wrote a story at 63–64 and code at 85–88 tokens/s with its default drafting, the GGUF runner 18–20 without; reading was timed on a 32K prompt against 4K. The last two are two different community machines, and the GGUF runner had one card more. Full conditions are in the GitHub folder.

    04 · What this does not show

    Estimates, checked against real runs

    • 8–35%median miss of the estimated writing speed per stack, on machines the fit had not seen; 35–85% for reading. One stack was measured on one machine only (128% for reading). The range beside each estimate covers about eight in ten such runs. All one stream; more streams are extrapolated
    • Not your workRuns use short questions and default settings, with nothing running on the tools card. Test your own documents, prompts and vision work before buying
    • Answer qualitySpeeds only. Quantising costs answer quality, more in the smaller versions: GLM’s 2-bit version scores 13% lower on one reasoning test. Check them on your own tasks
    • Prices moveListed retail prices on 15 November 2026, including GST. Freight and assembly are extra