The number that matters to me isn’t the 40 on the Intelligence Index, it’s the input side: 8 billion active parameters per token for reading context, and a KV cache about 75 percent smaller than V4 Flash. That’s a model shaped for the work I actually do — dumping a repo, a pile of tool outputs, and a million tokens of context at something and asking it to reason across all of it.
The catch is the chatty output. V4.1 Flash burns noticeably more output tokens per task than V4-Pro, so the cheap per-token price isn’t the whole bill. Artificial Analysis lands it at $0.27 per task, about 60 percent under V4-Pro — but V4 Flash did the same task for $0.22. Weights are MIT, so I’d measure my own token burn before switching anything.
The story — DeepSeek released V4.1 Flash, an open-weight MIT-licensed model with 552 billion parameters, native image input, and a one-million-token context window. A new causal encoder-decoder architecture splits input processing from generation, cutting compute and memory. V4 Flash is discontinued; from September 14, V4-Pro requests route to V4.1 Flash at its rates (Source).