DeepSeek V4.1-Flash: 763B model beats V4 Pro on AA Index 40

DeepSeek's release of V4.1-Flash follows a year in which the lab kept a lower profile, letting peers like GLM and Kimi lead on open models. The new model marks a deliberate architectural break from prior DeepSeek generations, introducing a causal encoder-decoder design that separates input and output computation. As Sebastian Raschka noted, the change is substantial enough that DeepSeek "should have called it DeepSeek V5." The naming choice—only a 0.1 version bump—may understate the shift.

The benchmark picture is mixed. V4.1-Flash sits behind some open models on the Artificial Analysis Intelligence Index, but the body argues this reflects benchmarks that don't yet capture what DeepSeek is optimizing for: efficient use of context. The model's reported KV cache footprint is up to 1/8 that of V4 Flash, and local deployment reports show it running on consumerish hardware at 200+ TPS. This suggests DeepSeek is prioritizing servability—the ability to run on offload pipelines and quantized KV stacks—as much as raw benchmark scores.

The stakes hinge on whether the model's verbosity undermines its low per-token pricing. Artificial Analysis measured 89k tokens per Intelligence Index task, 62% more than V4 Pro 0813. If that verbosity holds in real agent workloads, the total cost advantage may narrow. For inference engineers and teams running long-running agents, the question is whether V4.1-Flash's architecture translates to cheaper production deployments, not just cheap benchmark runs.