Imagine a laptop with 32 GB of memory and a 1 TB drive. Its owner sees a new AI model available to download and thinks: there is plenty of storage left, so why not run it here?
The catch is that storing a model and running it are different jobs. A large drive is a cupboard. Working memory is the table where the computer does its work. Buying a bigger cupboard does not automatically enlarge the table.
Mistral’s October 29 release makes this distinction unusually vivid. The company has released the weights of Mistral Large 4, also called Le Chonk. You can obtain the model itself instead of relying entirely on Mistral to answer requests. But its model card lists roughly one trillion parameters. That number deserves a pause before the download.
First, count the numbers you need to store
Weights are the numerical settings learned during training. They help determine how a model transforms an input into an output. Those numbers occupy space, and the amount depends on how each number is represented.
Here is a deliberately simplified calculation for exactly one trillion stored values. A byte contains eight bits; GB below means one billion bytes.
| Bits per value | Calculation | Space |
|---|---|---|
| 16 | 1 trillion × 2 bytes | 2,000 GB |
| 8 | 1 trillion × 1 byte | 1,000 GB |
| 4 | 1 trillion × ½ byte | 500 GB |
These are arithmetic examples, not measured download sizes or a shopping list for hardware. Real packages can mix number formats and include extra information. Running the model needs additional memory too.
Still, the scale is informative. Even the hypothetical 500 GB version cannot sit entirely inside our imagined laptop’s 32 GB of memory. Some other arrangement would be necessary. Having enough disk space answers only the first question.
How smaller numbers help
Quantization reduces the precision used to represent weights. Think of recording a measurement with fewer digits: you save space, but you must check whether the lost detail matters for the job.
Hugging Face’s October 29 release guide lists official FP8 and NVFP4 variants. The names describe numerical formats, not a promise that every computer can run them. Mistral’s NVFP4 card says that version targets NVIDIA Blackwell GPUs and mixes formats across different parts of the model. Our neat four-bit calculation is therefore not its actual file size.
A smaller representation can make a model more practical. Whether it preserves the quality you need, and whether your software and hardware support it well, remain separate checks.
Why “52 billion active” does not mean a 52-billion-parameter download
Large 4 uses a design called a mixture of experts. According to Mistral, about 52 billion parameters are active per token when embeddings and output layers are included, out of roughly one trillion total. A token is a small unit of text, often part of a word.
In this design, a learned router directs a token through selected expert networks. “Expert” means a part of the neural network; it does not necessarily mean a named specialist in chemistry or accounting. Hugging Face’s MoE explanation describes the resulting distinction between computation and memory.
Picture a library where a reader consults a few books at a time. Fewer books open on the desk means less reading at that moment. The other books still belong to the library. Likewise, unused experts do not vanish from the model because a particular token did not need them.
The active count helps describe how much of the model participates in a step. The total count helps explain why the complete collection of weights can still be enormous.
The conversation needs room as well
A running model also keeps temporary information. One example is the key-value cache, which saves calculations from earlier tokens so they can be reused while generating later ones. Transformers’ cache guide explains how its memory use changes with the cache strategy and model design.
A longer conversation can require more cache space. That is one reason “the weights fit” is an incomplete answer to “will this workload fit?” The advertised maximum context length also does not tell you how much conversation your particular setup can handle comfortably.
There are ways to spread the load. Accelerate documents offloading weights between GPU memory, ordinary computer memory and disk. This can make otherwise oversized models runnable, but moving data introduces overhead. It establishes a technique, not evidence that Large 4 will be pleasant to use on our hypothetical laptop.
There is a concrete middle ground between an ordinary laptop and a server installation. The llama.cpp team lists a compressed variant at 298 GB and says it fits in 512 GB of unified memory. That is a documented option for a very large-memory system, not a performance test conducted for this article.
What open weights change for you
Open weights give people and organizations more choices about who runs a model and how it is adapted. Computing still has a cost. The release license also has conditions; downloadable does not mean every use is unrestricted.
For someone who wants to summarize a few notes locally, a smaller model may be the more useful starting point. For scale, eight billion values at four bits each occupy an idealized 4 GB before extras. That arithmetic does not establish answer quality, but it explains why model size changes the range of plausible devices.
Start with the task: a short summary, a long document or a coding session. Then check the exact model variant, its supported hardware, the memory required for that workload and the speed you find usable. The most useful model for your computer is the one that handles your work well enough within those limits.
