Nota cuts Kimi K3.1 Pro cost by running it on B200 GPUs
Nota pruned up to 37% of the model's expert modules and says the optimized Kimi K3.1 Pro keeps over 99% of its GPQA Diamond score.
Nota said on November 10, 2026 that it has optimized Moonshot AI's Kimi K3.1 Pro so the model runs on eight Nvidia B200 GPUs instead of eight Nvidia B300 GPUs, according to datanet.co.kr. The company said the switch cuts initial GPU purchase costs by up to about 20%, based on its own quote.
The optimization relies on pruning that removed up to 37% of the model's expert modules to lower memory requirements. Nota said the optimized model retained more than 99% of the original model's score on the GPQA Diamond scientific reasoning benchmark.
Context length and serving
On eight B200 GPUs, the version with 30% of experts pruned handles contexts up to 512,000 tokens, while the version with 37% pruned handles up to 1 million tokens. Nota said speculative decoding with a draft model is supported to speed up answer generation on the same eight-B200 configuration.
The company also said cloud rental quotes for B200 GPUs were up to about 22% lower than for B300 GPUs under the same model conditions. Kimi K3.1 Pro has 2.8 trillion parameters and is described as the largest open-weight AI model currently released by total parameters.
Previously, Nota said it had reduced the B300 GPUs needed to run the earlier Kimi K3 from eight to as few as four.
Quick answers
How many GPUs does the optimized Kimi K3.1 Pro need?
Nota said the optimized model runs on eight Nvidia B200 GPUs instead of eight Nvidia B300 GPUs.
How much does the optimization cut GPU costs?
Nota said initial GPU purchase costs fall by up to about 20%, based on its own quote, and that B200 cloud rental quotes were up to about 22% lower than B300 quotes under the same model conditions.
Does pruning hurt the model's benchmark score?
Nota said the optimized model retained more than 99% of the original model's score on GPQA Diamond.