Build1 publisher2 min readPublished Updated
DeepSeek's new encoder-decoder splits inference into an 8B prefill and a 16B decode
V4.1-Flash retires the V4 Pro line and carries two active-parameter counts, 763B total with 8B on input tokens and 16B on output, so one sizing number no longer covers both phases of a request. Baseten had it running on day zero.
The Engineer · Build desk
What happened
- DeepSeek released V4.1-Flash with 763B total parameters, 8B active on input tokens and 16B active on output, a 1M-token context, text and image input, and an MIT license, per Artificial Analysis.
- Latent Space reported that DeepSeek is retiring V4 Pro and going all in on a completely new causal encoder-decoder architecture, released under a point-release version number.
- Artificial Analysis scored the model 40 on its Intelligence Index and said it surpasses DeepSeek V4 Pro 0813 while costing much less to run.
- Artificial Analysis listed the price at $0.30 per million input tokens and $1.20 per million output, with cached input at $0.006 and a further 50 percent off-peak discount.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Sizing a deployment now takes two active-parameter counts, since decode activates twice what prefill does, and a batching profile tuned to a single figure will mis-size one of the two phases.
- decision Cached input is 50 times cheaper than fresh input, so cache retention policy sets what an agent costs to run.
- cost Generation-heavy agents pay twice for the same tokens: output is priced at four times input, and it runs the heavier 16B path.
- contradiction Artificial Analysis places the model below GLM-5.3-Flash while Vals ranks it first among open-weight models ahead of Kimi K3, so the index a team standardizes on decides whether this looks like a frontier buy or a cheap runner-up.
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 [4]. 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 [5]. A capacity planner that stores one active-parameter count per model is now storing the wrong thing, and decode is twice prefill [15].
Against the 763B total, 8B is about 1.0 percent and 16B about 2.1 percent [16]. Latent Space put the sparsity at 1 to 2 percent [6].
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 [7]. At an eighth of the footprint, the same memory budget holds eight times as many tokens [20]. The price list points the same way: cached input is 50 times cheaper than fresh input [18], and output is four times input [17]. Off peak, both halve, to $0.15 and $0.60 per million tokens [19].
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 [11][10]. 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" [13], and argued the target is "the most creative and efficient use of context we have ever seen openly explained" [14].
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 [12]. 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 [2].
What to watch
- A per-phase memory breakdown from DeepSeek would let operators size prefill and decode hosts without measuring them first.
- Whether serving stacks beyond Baseten ship the encoder-decoder path, and what they report having to change for it.
- A Vals rerun with an output cap above 384 tokens would show what the 16B decode path costs per test in practice.