📄 中文摘要
Dan Woods的实验展示了流式专家技术的应用，该技术允许在内存不足以容纳整个模型时，通过从SSD流式传输必要的专家权重来运行更大的混合专家模型。最近，Dan在48GB内存的情况下成功运行了Qwen3.5-397B-A17B模型。而在短短五天后，另一位用户在96GB内存的M2 Max MacBook Pro上成功运行了Kimi K2.5模型，该模型拥有1万亿参数，并且在任意时刻激活32B权重。这一进展展示了流式专家技术在处理超大模型方面的潜力。
流式专家
发布: 2026年3月24日
📄 English Summary
Streaming experts
Dan Woods' experiments have showcased the application of streaming experts technology, which allows for running larger Mixture-of-Experts models on hardware that lacks sufficient RAM by streaming the necessary expert weights from SSD. Recently, Dan successfully ran the Qwen3.5-397B-A17B model with 48GB of RAM. Just five days later, another user reported running the colossal Kimi K2.5 model, which has 1 trillion parameters and 32B active weights at any given time, on a 96GB M2 Max MacBook Pro. This advancement highlights the potential of streaming experts technology in handling ultra-large models.