The interesting thing about DeepSeek V4 isn't the benchmark. It's the release engineering.

When DeepSeek dropped the V4 family on April 24 — MIT-licensed, weights on the Hub, 1.6T total / 49B active for Pro and 284B / 13B for Flash, both with a million-token context — the model didn't arrive alone. vLLM and SGLang each published an official serving recipe the same day. Not "support landed a week later." Day zero. Native CSA+HCA attention, FP4 MoE backends, MTP speculative decoding, disaggregated prefill/decode — all wired up and documented before most people had finished downloading the checkpoint.

That coordination is the story, and it changes how you should think about adopting an open model.

The license is not the thing you're adopting#

Here is the part nobody puts in the launch tweet: the checkpoint is FP4+FP8 mixed. The MoE expert weights — the overwhelming bulk of the 1.6T — are stored in FP4 (and not all FP4 is the same format). The attention, norms, and router stay in FP8. That is not an incidental quantization choice you can undo. It is how the model was trained to be served, and it means the throughput you were promised is only real on hardware that has FP4 tensor cores and kernels written for them.

Read the day-0 recipe and you find the actual dependency list. Peak performance leans on FlashMLA, DeepGEMM's Mega-MoE kernels, and the TRTLLM-Gen fused MoE backend routed through FlashInfer — a path that pairs MXFP8 activations with MXFP4 expert weights and targets Blackwell specifically. The PyTorch/SGLang writeup reports roughly 5× higher throughput at the same interactivity on GB300 versus the launch baseline. That number is not a property of the weights. It is a property of the kernels.

An open-weight model's real license is which GPU it runs fast on.

So the model is MIT and portable in the sense that you can legally run it anywhere. But "portable" and "economical" have quietly diverged. You can serve V4 on your A100s. You will not get the tokens-per-dollar the launch charts implied, because the charts were drawn on Blackwell with kernels that don't exist for your card. The weights transferred. The performance didn't come with them.

The clearest proof: the model got faster while the model stayed the same#

SemiAnalysis tracked the same DeepSeek V4 checkpoint from day 0 to day 43 across GB300 NVL72, B200, MI355X, and Huawei silicon. Throughput climbed steadily over those six weeks.

Sit with that. The weights were frozen the whole time. Nobody retrained anything. What improved was the serving stack underneath — better kernels, better expert-parallel load balancing, tighter prefill/decode disaggregation. The "day 0 to day 43" curve is the cleanest illustration you'll find of a truth that open-weight releases keep obscuring: the checkpoint is a constant, and the inference engine is the variable. When you pick a model in 2026, roughly half of what you're actually choosing is the maturity of the runtime that serves it.

This is why the labs now co-ship the recipe. A model whose day-0 serving story is a mess reads as slow, and "slow" gets attributed to the weights even when it's the kernels. DeepSeek learned the lesson and treated the vLLM and SGLang recipes as part of the launch, not a downstream community chore. The serving engine has been absorbed into the release.

What this means when you're the one deploying it#

Three practical consequences, none of which are about the benchmark:

DeepSeek gave away a frontier-class model for the cost of a git clone. That's genuinely remarkable, and it's also the misdirection. The weights were the easy part to open. The thing that's still, functionally, closed — gated behind specific silicon and the kernels written for it — is the speed. Read the recipe, not the license.