Bleys 算了一笔账：GPT-5.6 Sol 参数 2-4 万亿、跑在 Cerebras 晶圆上、速度 750 token/s，但价格可能翻倍。想了解下一代超大规模模型怎么造？看这个。
Bleys 根据公开信息估算，OpenAI 的 GPT-5.6 Sol 模型总参数在 2-4 万亿之间，采用 MoE 架构，每任务激活约 1500 亿参数。模型被分布在 70-100 个 Cerebras 晶圆上，每晶圆是一个巨型芯片，全模型约 70-90 层，最多每晶圆一层以避免瓶颈。由于专为 Cerebras 芯片设计，推理速度可达 750 tokens/s，但使用成本可能比现有模型高出数倍。
Bleys estimates GPT-5.6 Sol is somewhere between 2…
Bleys estimates GPT-5.6 Sol is somewhere between 2-4 trillion parameters total.
He calculates they're splitting it across 70-100 Cerebras wafers (each wafer is one giant chip). The model likely has around 70-90 layers. At most one layer per wafer so everything runs without bottlenecks.
Probably Mixture-of-Experts. 2-4 trillion total parameters, but only around 150 billion activate for any single task.
He thinks they designed the model specifically for these Cerebras chips rather than using older approaches. That's how they're getting to 750 tokens per second.
It will probably cost more to use. Bleys estimates several times the usual price because of the hardware setup.
OpenAI is building models alongside hardware companies now. Not just raw size. Power and speed.
The image shows Cerebras' calculator with these numbers plugged in to show the claimed performance.