According to CoinMeta, the inference engine WASTE has been open-sourced and is capable of running Kimi K3 seamlessly on a 64GB MacBook Pro device. The transformed model occupies approximately 1TB of space and processes token at a rate of 0.49 to 0.54 per second. WASTE retains about 27GB of the model's core structure in memory, with over 80,000 experts stored within its built-in SSD. Only the experts currently in use are fetched from the hard drive when each token is generated. This version does not involve distillation, pruning, or deletion of experts; however, the weights of these experts have been re-quantized to a range of 3-, while the model's core structure uses 4- and 8- bit formats.