Open-source video models

Member lesson

The hardware you really need

Lesson 6 of 10 in Open-source video models.

Checked against the makers' documentation on 25 September 2026.

Optimize Your AI - Quantization Explained · Matt Williams
Lesson

The first sum is simple. A model needs roughly its parameter count times the bytes per parameter just to hold its weights, before any memory for the clip it is making. Wan 2.2 TI2V-5B shows the whole ladder in files you can download today, since its transformer is 20 GB in the 32-bit Diffusers repository, 10.0 GB as Comfy's 16-bit file, 5.4 GB as a Q8 GGUF, 3.43 GB at Q4_K_M and 1.85 GB at Q2_K. Each step down fits a smaller card and gives up some quality, and the text encoder and VAE need room on top.

32-bit · 20 gb16-bit · 10 gbgguf q8 · 5.4 gbgguf q4 · 3.4 gbgguf q2 · 1.9 gb8 gb card24 gb cardweights only: the encoder and the clip need room on top
The Wan 2.2 5B memory ladder: 20 GB at 32-bit, 10 GB at 16-bit, 5.4 GB at GGUF Q8, 3.4 GB at Q4 and 1.9 GB at Q2, against 8 GB and 24 GB graphics cards; weights only.
An ink drawing in three panels: a green sofa in a room, the same sofa flat-packed into three boxes, and the boxes squeezed into the open boot of a small red car.
Quantising and offloading are packing: the sofa fits the small car, and something is given up to get it in.
Member

The rest of this is for members.

You have read the opening of this lesson. The rest of it comes with a membership.

One membership covers all four courses and Brands.

  • All lessons in the four Learn courses
  • All brand breakdowns in the Brands section
  • All quizzes, flashcards and crosswords
  • An AI teacher on every lesson, for up to 60 questions a day
  • All course material published during the membership

Price

Membership is from $12 a month on the Yearly plan, which is charged at $144 once a year. Lifetime access is $59, paid once.Membership is from ₹200 a month on the Yearly plan, which is charged at ₹2,400 + GST once a year. Lifetime access is ₹3,999 + GST, paid once.

Check what stuck

Members only

You can't take this check yet.

Members answer these 3 questions and keep the ticket for this lesson. Every lesson has one.

  1. Question 1, locked

    Which of these words means this? “The memory on the graphics card itself, where the weights and the working space for a clip have to fit or be shuffled in and out.”

    • Parameters
    • Pipeline
    • VRAM
    • Node
  2. Question 2, locked

    Which of these words means this? “How many bits store each number; 32-bit floats take four bytes per parameter, bfloat16 two and fp8 one.”

    • Precision
    • VAE
    • Custom node
    • Community licence
  3. Question 3, locked

    Which of these words means this? “Storing weights in fewer bits, as fp8, int8 or a GGUF level such as Q4_K_M, so they take less memory at some cost in quality.”

    • Quantisation
    • Parameters
    • Pipeline
    • Node

Elsewhere in the library

Hear about new courses.

Leave an email and we will tell you when something new lands. We do not write often.