TOOL VERDICTS

Can You Run a 1T Model on Your Own Iron?

Mistral's 1T-parameter 'Le Chonk' is Europe's biggest open-weight release ever. We did the hardware math: 1.05T params needs 2.1TB at FP16, about 1TB at FP8. Your 8xH100 box has 640GB.

2026-10-07

Mistral Large 4, “Le Chonk,” dropped Oct 6. 1.05 trillion parameters, open weights promised. So we asked the only question that matters: can you actually run it?

The math: 1.05T params needs 2.1TB at FP16, about 1TB at FP8. Your 8xH100 box has 640GB. This is a sovereign-AI play, not a homelab model.

The full VRAM breakdown first ran as an Instagram carousel, with a reel twin on the same day. Two catches nobody mentions: the weights are not out yet (end of October), and the license is still unpublished.

The spec

1.05T params. Only 49B awake per token.

Le Chonk is a Mixture-of-Experts model. A router wakes a small crew of expert networks per token, so compute per token looks like a 49B model. But all 1.05T parameters still have to live in memory. Active compute is small. Resident memory is not.

The math

FP16 needs 2.1TB. FP8 needs 1TB.

1.05 trillion params times 2 bytes is 2.1TB. At FP8, 1 byte each, it is 1.05TB. That is weights alone, before a single token of KV cache. Quantization shrinks the weights. It does not shrink the problem to something a single box can hold.

The box check

Your 8xH100 box has 640GB. Not enough.

8 times 80GB is 640GB. FP8 weights alone need about 1,050GB, so eight H100s are still 410GB short of loading Mistral Large 4. The realistic floor is 16xH100 (1.28TB) or two 8xH200 nodes, before KV cache for the 1M-token context window.

The catch

Nobody can self-host it yet.

Weights land end of October (Hugging Face lists Oct 31). The license is unpublished. Large 3 was Apache 2.0; this one is expected to carry a custom Mistral license. Until then it is API only: $1.36 in, $4.18 out per 1M tokens.

The claim check

“Best open model.” With an asterisk.

On the Artificial Analysis Intelligence Index, Le Chonk scores 38, the best outside the US and China. But seven Chinese open models score higher, topped by MiMo-V2.6-Pro at 46. The regional claim has one independent index behind it. The global claim does not hold.

The verdict

Sovereign play. Not a homelab model.

Save this VRAM sheet before you size a cluster:

Precision Weights Minimum hardware
FP16 2.10TB Two 8xH200 nodes, minimum
FP8 1.05TB 16xH100 (1.28TB) or two 8xH200
FP4 0.53TB 8xH100 fits the weights; the KV cache will not

Your move, depending on who you are:

  1. Test it today: the API runs $1.36 in and $4.18 out per 1M tokens.
  2. Sizing a cluster later: save the VRAM sheet above.
  3. Before you commit: read the license when it drops.

What we did not verify

This was hardware arithmetic, not a run of the model. Nobody ran anything, because the weights were not out yet. The numbers above are multiplication on published specs, not a benchmark.

The license situation is an expectation, not a fact: Large 3 was Apache 2.0, and a custom Mistral license is expected for Large 4, but nothing was published at write time.

The “best open model” check rests on a single independent index, the Artificial Analysis Intelligence Index. And the caption’s original claim, “biggest open-weight release of the year,” was wrong: Kimi K3 (2.8T, weights released) is bigger. The corrected claim is “Europe’s biggest open-weight release ever.”

The voiceover in the reel twin is AI-generated. Watch the reel on Instagram or the YouTube Short.

This post started as an Instagram post →

Discussion

Talk it through

Argue with us on Instagram.