DeepSeek's weights silently swapped on Vercel AI Gateway
Both V4 Pro and Flash now serve new weights by default — live requests changed with no code edit. Pin the dated snapshot before you trust last week's evals.
Copy markdown
Update: V4 Pro's 0813 build is now the Gateway default
On Aug 12, Vercel AI Gateway made deepseek-v4-pro-0813 the default behind deepseek/deepseek-v4-pro, so live traffic picked up the new weights with zero code change. Pin deepseek-v4-pro-0813 in the AI SDK if you need reproducible output — the Gateway takes no markup and honors BYOK, so a dated snapshot is free insurance.
V4 Flash swapped too — re-run your evals
DeepSeek V4 Flash got the same silent weight update on the Gateway the same day. If you benchmarked or shipped against last week's build, outputs can drift; audit every model you left unpinned across your gateways before trusting old numbers.