DeepSeek's new encoder-decoder splits inference into an 8B prefill and a 16B decode

Send a request and the two halves of the model bill differently. Input tokens run the 8B active path, and every output token runs the 16B one ^[1]. To write that down, Latent Space extended the established naming convention for MoEs into 763B-P8B-D16B, because the usual notation carries one active count ^[2]. A capacity planner that stores one active-parameter count per model is now storing the wrong thing, and decode is twice prefill ^[10].

Against the 763B total, 8B is about 1.0 percent and 16B about 2.1 percent ^[20]. Latent Space put the sparsity at 1 to 2 percent ^[19].

The cache claim is the one to check first if you run long agents. Reading the v4.1 Flash tech report, Latent Space credits tweaks including Sliding-Window Attention Bounded Replay ^[4]. At an eighth of the footprint, the same memory budget holds eight times as many tokens ^[14]. The price list points the same way: cached input is 50 times cheaper than fresh input ^[12], and output is four times input ^[11]. Off peak, both halve, to $0.15 and $0.60 per million tokens ^[13].

Vals ran its index at 1M context with a 384-token output cap, temperature 1, default top-p and top-k, and high reasoning effort, and reported $0.30 per test ^[9][8]. A 384-token cap does not exercise the 16B decode path for long. The per-test figure transfers to workloads whose outputs stop around there; a coding agent that streams a few thousand tokens a turn buys the 16B path many more times per turn.

Latent Space wrote that v4.1 Flash "is technically behind other open models in some benchmarks" ^[7], and argued the target is "the most creative and efficient use of context we have ever seen openly explained" ^[16].

On adoption, the recap has one line: Baseten shipped day-0 support and described the model as smarter, faster and more efficient than DeepSeek v4 ^[15]. The recap does not say what that support cost to build. A new architecture with two active counts and an image input path is more than a weight swap for a serving stack, and Baseten going live on day zero is the only measure of that work here. The model is MIT licensed and served first-party by DeepSeek in the US, according to Artificial Analysis ^[17].