Local AI Advisor
All posts

Mistral Large 4 ('Le Chonk'): why a 1-trillion-parameter open model still won't fit on your PC

France's Mistral just previewed a 1-trillion-parameter open-weight model. The headline number is exciting; the memory math says your home PC can't run it.

By Ray, Your Local AI Advisor4 min readlocal-llmannouncementsbeginnersnew-modelshardware

France's Mistral AI announced Mistral Large 4 — nicknamed "Le Chonk" by the company itself — on October 6, 2026. It is a natively multimodal, open-weight model with 1.05 trillion total parameters and 49 billion active per token, per Mistral's own documentation. The model is live now only as an API preview; Mistral says the weights ship by the end of October. So the question for this blog: when those weights land, could you actually run it on your own machine?

The honest answer is no — and the math behind it is a skill worth learning, because it lets you judge every future giant model the same way.

Flat illustration of a giant robot-server towering over a tiny laptop on a desk, showing the scale gap between a trillion-parameter model and a home PC

Total vs. active: the number that actually matters

Large 4 is a mixture-of-experts (MoE) model. Think of it as a building with many specialist offices, but only a few offices' lights are on for any single task. Mistral reports 1.05 trillion total parameters, but only 49 billion active per token — roughly 1 in 20 parts of the network wake up for each word it writes.

Diagram showing a grid of gray cubes with only a few glowing orange, connected by arrows to an output node, illustrating mixture-of-experts

That active number tells you how much compute each answer costs — roughly what a 49B dense model would need per token. But here is the catch that marketing glosses over: every parameter still has to live in memory, active or not. The lights-off offices still take up floor space. So the memory bill is set by the 1.05 trillion, not the 49 billion.

The honest memory math

The rule of thumb: a model needs about half a byte of memory per parameter at 4-bit quantization, plus a little overhead — call it 0.55 bytes per parameter for a safe estimate. At 8-bit, double it. That estimate covers only the weights; the KV cache and runtime add more.

  • Mistral Large 4 (1.05T): ~578 GB at 4-bit, ~1.15 TB at 8-bit.
  • For comparison: Reflection Beam (501B) ~276 GB, DeepSeek V4 Pro (1.65T) ~908 GB, Qwen3.8 Max (2.4T) ~1,320 GB — all at 4-bit, weights alone.

Horizontal bar chart comparing estimated 4-bit weight sizes: Qwen3.8 Max about 1320 GB, DeepSeek V4 Pro about 908 GB, Mistral Large 4 about 578 GB, Reflection Beam about 276 GB, with an RTX 4090 24 GB line

The red line in the chart is an RTX 4090 with 24 GB. A fully loaded M4 Max machine with 128 GB of unified memory — the biggest consumer setup most readers own — holds less than a quarter of Le Chonk's 4-bit weights. Serving this model takes a multi-GPU server: eight 80 GB GPUs give 640 GB, which barely fits the 4-bit weights before the KV cache eats the rest. Until community quants or distills appear, this is a datacenter model.

What this changes for you today

Nothing on your machine — but three things worth doing:

  1. Use the same math on every announcement. When the next "X trillion parameters" headline drops, multiply by 0.55 bytes for 4-bit. If the answer dwarfs your hardware, it's a server model, full stop. The active-parameter count tells you how fast it thinks; the total count tells you how much room it takes.
  2. Watch the distills, not the giant. Big labs now routinely release smaller distilled versions of frontier models — a future 49B-class Le Chonk distill would need roughly 28 GB at 4-bit, which fits a 32 GB GPU or a 64 GB unified-memory Mac. That's the tier that actually reaches your desk.
  3. Try it as an API today. Mistral's public preview is live on its platform, so you can judge the quality for yourself while the weights are pending — just know an API trial tells you nothing about local performance.

Also keep your eyes on Mistral's Hugging Face organization: as of the announcement, no Large 4 checkpoint had appeared there, so the weights drop is a "watch, don't wait" situation. And remember that Mistral's benchmark claims — strongest open model outside China "by a substantial margin," 93% on Cybench — are the company's own numbers. Independent evaluations will land after the weights do, and those are the ones worth trusting.

Don't plan around it yet

Le Chonk is a schedule, not a download. The weights are promised by end of October under an open-weights release, but the exact license terms come with the weights — Mistral's previous open release, Large 3, shipped under Apache 2.0, but nothing is confirmed for Large 4 until the files land. Don't reorganize your local setup or your budget around a model you can't pull yet. If you want to know what your current machine can run today, use /check and the /learn hub instead — the giant models will still be there when they're actually downloadable.

Sources