Content
@
0 reply
20 recasts
20 reactions
Buldevskaya
@buldevskaya
Nvidia just dropped their heavyweight reasoner. It’s built on Llama 3.1 405B, fine-tuned down to 253B parameters. Outperforms DeepSeek R1 across almost all benchmarks — unlike QwQ 32B, which only led in a few. The model runs FP8 on 4×H100s, but if you’re going BF16, you’ll need 8×H100s or 4×B100s. For us mortals, they’ve also released smaller versions a couple of weeks ago — 8B and 49B, based on Llama 8B and 70B.
0 reply
0 recast
1 reaction
