Will it fit?
Le Chonk in fp8, one 32K chat: 1.08 TB.
8 × H200: not enough. It takes 9.
At 4 bits: 5 H200s, 4 B200s, or 2 Mac Studios with 512 GB each.
$CHONK
Rent the GPUs or pay the API?
One person on rented H200s: $143 per million tokens written.
Mistral's API: $8.26 (with 3 tokens read per token written).
64 people sharing 8 B200s: $7.29. Just.
$CHONK
Before you download a terabyte, ask it:
chonk fit mistral-large-4 --hw h200 --count 8
"It does not fit on 8 x H200 141 GB: it needs at least 9."
Open source, one file, 13 tests, runs right on our site. $CHONK