A series · signalandweight.com

Signal & Weight

Where model weights live, how they move, and what they cost: one real model followed from storage to GPU memory to the next token.

Every part follows the same real model, Qwen3.8-Max, so each number connects to the ones before it. Each part comes back to at least one figure you can check yourself from the model's published config.

• Part 1

  One token through a 2.4-trillion-parameter model

  Every word an LLM writes costs a full trip through its weights. I followed one prompt through Qwen3.8-Max to see what that trip actually looks like, in bytes, GPUs and memory.

• Part 2

  What weights actually store: compression, not a database

  Coming soon

• Part 3

  Quantization: the dial that decides how many GPUs you need

  Coming soon

• Part 4

  Weights at rest: files, shards and formats

  Coming soon

• Part 5

  Getting weights onto GPUs: the cold-start problem

  Coming soon

• Part 6

  The KV cache: the memory that is not weights

  Coming soon

• Part 7

  Mixture of Experts at scale: why batching touches every expert

  Coming soon

• Part 8

  Training checkpoints: seven times bigger, and the restore is what hurts

  Coming soon