DeepSeek open-sources its engine as Washington moves to ban it

DeepSeek released the inference framework behind its V4.1 Flash service, with claims of high-throughput, low-latency serving at large-cluster scale. The top comment called it "a nuclear bomb" and another said US frontier labs have no path to trillion-dollar valuations from here. Within hours the DHS order banning DeepSeek from federal civilian agencies was on the front page too. One commenter's reply summed up the ban: V4.1 Flash is "out of this world good ... and cheap, that's why."

The pattern here: the model weights were already open, now the serving stack is too. That removes the last moat Western labs claimed, which was that they alone knew how to run this stuff cheaply at scale. The policy response is going the other way. Commenters pointed out the irony that open weights can be inspected for malware, while US inference providers can't be. MiniMax M3.1 Flash showed up the same day with a claimed 3x inference speedup, and people called it another step up the price/performance frontier.

The counterpoint is that a federal civilian ban is narrow and a lot of the thread treated it as geopolitics, not security. But a ban at the federal level tends to spread to contractors and regulated industries.