DeepSeek V4 Flash now runs updated weights on AI Gateway

DeepSeek V4 Flash now runs updated weights by default on AI Gateway. The change is automatic for traffic to deepseek/deepseek-v4-flash: the same model ID and existing request code pick up the new weights with no migration step. The update is framed around stronger agentic capabilities rather than a rename or new product surface.

On Terminal-Bench, the updated model scores 82.7. That is 25.8 points above the April preview result of 56.9. The benchmark jump is the concrete performance signal tied to the weight update; AI Gateway surfaces it by routing default deepseek/deepseek-v4-flash calls to those weights without a separate endpoint or version string change.

For engineers already calling DeepSeek V4 Flash through AI Gateway, the practical effect is an in-place capability lift on agent-style workloads. Terminal-Bench-oriented loops, tool use, and multi-step task completion can improve without client changes, redeploys, or ID swaps. That reduces the usual cost of chasing a better checkpoint when the provider controls default weights behind a stable model string.

On the supply side, DeepSeek is currently the only provider serving the updated weights. Other providers have not yet matched that serving path, so AI Gateway traffic that lands on DeepSeek is the channel that receives the Terminal-Bench gains today. That is a temporary exclusivity window: builders comparing multi-provider routes for the same model ID may see uneven quality until others ship the same weights.

What to watch next is when additional providers start serving the updated DeepSeek V4 Flash weights and whether Terminal-Bench and agentic quality stay aligned across them. Until then, treat deepseek/deepseek-v4-flash on AI Gateway as the path that already defaults to the stronger weights, and re-check provider routing if you pin or load-balance across non-DeepSeek backends that may still be on the older checkpoint.