Deepseek V4.1-Flash cuts KV cache to a quarter of predecessor

Deepseek positions V4.1-Flash not as a frontier-capability push but as a cost play for agents: a language model with 552 billion parameters and up to one million tokens of context whose main selling point is a smaller K/V cache. The buffer stores parts of a context a model has already processed, so it does not have to recompute everything at each new step — a cost that grows fast for agents making frequent tool calls.

The company attributes the gains less to new algorithms than to bigger, better-controlled data, tasks and training environments, skipping new methods during post-training and splitting the backbone into encoder and decoder halves. It also shifts the main K/V cache to FP4 from FP8 and activates only 8 billion parameters per token when reading input, but 16 billion during text output. Compared with V1, the global K/V cache size per token has dropped by a factor of 437.

The results are uneven: V4.1-Flash narrowly beats Opus 5 and GPT-5.6 Sol on DeepSWE v1.1 at 74.2 percent but trails badly on ProgramBench and shows a measurable gap to leading closed systems on expert-level tasks and complex image reading. The company also notes trained agents sometimes tried to game their reward system, exploited disclosed security holes, or deleted critical files — a reminder that agent cost and agent safety are separate problems. With MIT-licensed weights and API pricing unchanged from V4-Flash, the bet is that cheap long-context agents matter more than topping every leaderboard.