Cerebras 跑 V4.1 Flash 或只需不到 10 片晶圆

GDP @bookwormengr原推发布于 2026.10.23

在 X 上打开原推

True, Cerebras needs 60 wafers for K3. That is because K3 is huge model with 2.8 trillion parameters. With 1.5TB in storage. But new architectures allow it to cut down that count to just 10 with total hardware cost approaching a single NVL72, imho. Most near Mythos capable models for cyber are around 800B (GLM 5.3, DeepSeek v4.1 flash) Moreover for a model like DeepSeek v4.1 flash, you need only 500B weights on wafer; remaining 192B are embeddings can be hosted off chip. This brings down wafer need drastically to less than 10!!! Moreover; prefill needs only half the layers in DS architecture that is going to be commonplace in coming days and KV is quite small; so Cerebras can also greatly perform prefill (something they avoid) and do decode with larger micro batches. It will likely be quite efficient with just 10 wafers; and will do well for Cyber, Design and voice use cases, imho.

这条是引用着另一条写的：x.com/joefioti/status/2113344525121003579