{"type":"rich","version":"1.0","title":"deadmanoz wrote","author_name":"deadmanoz (npub12m…6pvdp)","author_url":"https://yabu.me/npub12msugd9s2xz5p6flnm9ygsx2zcgnnr29tw0ezq53z3p06xag70vqv6pvdp","provider_name":"njump","provider_url":"https://yabu.me","html":"“The 1.8-bit (UD-TQ1_0) quant will run on a single 24GB GPU if you offload all MoE layers to system RAM (or a fast SSD). With ~256GB RAM, expect ~10 tokens/s. The full Kimi K2.5 model is 630GB and typically requires at least 4× H200 GPUs.\n\nIf the model fits, you will get \u003e40 tokens/s when using a B200.”\n\nhttps://unsloth.ai/docs/models/kimi-k2.5"}