What exactly is the cheap big bowl model? Tactical leaning back. The gospel of local deployment of large models has arrived! Here’s a technical analysis of the recently released Kimi-Linear-48B-A3B from the Dark Side of the Moon! First, in one sentence — this is the cheap big bowl fast food model. The 48B-A3B level achieves around 1M context, and it uses linear attention, which is very memory efficient. The memory consumption from the growth of context length in traditional attention is exponential, while this one is linear, so this model can even run on a CPU without issues. I’m already downloading it, preparing to add it to my local commonly used models. Currently, the biggest uncertainty is not knowing how the recall level is. I plan to download it and feed it a few novels to ask about the details of the novels and see how the model responds to evaluate the recall effect. For those interested in the results, please give a thumbs up; if it exceeds 100, I’ll release the evaluation this weekend.