Leverage in a renewal negotiation is a function of what the alternative costs to stand up, and permissive terms are only the first line of that estimate. The line that usually decides it is serving efficiency, which is why the packaging here is worth reading closely. The SGLang instructions tell operators to enable DSpark without setting a separate draft model path, because the draft and target weights come from the same checkpoint [6]. For a platform team, that is one artifact to version and monitor rather than two [16], and the drafting mechanism itself is a documented flag rather than a research exercise: seven speculative tokens with greedy draft sampling in the vLLM example [5].
The board-deck version of this is that an open-weight model is now broadly competitive with the strongest proprietary models, so the API bill can come down. DeepSeek does make that claim on the model card [3]. It is incomplete for two reasons. The claim is the publisher's own, and part of the coding evidence rests on DSBench-FullStack and DSBench-Hard, which the card identifies as internal test sets [12]. The public code-agent results were produced inside DeepSeek's own harness in minimal mode at the max reasoning effort level, at temperature 1.0 and top_p 0.95 [11], a configuration a buyer would have to reproduce before quoting anything from it. The card text supplied to us refers to a benchmark table without figures we can read [15], and on comparative performance that leaves the claim unverified from the outside.
What can be sized today is the cost of holding the option open. DeepSeek's own example command serves the model on a single node of four GB300s with an fp8 KV cache, block size 256 and expert parallelism enabled [10], and it recommends allowing up to 384K output tokens at the high and max reasoning effort settings [8][9]. Generations at that ceiling are capacity planning, not a rounding error, and the shape of that launch command describes the skill profile a team needs on staff before the alternative is real.
Adoption is easier to read than performance. The model page reports 111,121 downloads in the past month [13], which works out to roughly 3,700 a day [14]. Those are pulls, not production deployments, so the number measures interest, and interest is not what a procurement officer can put in a slide.
The incumbent will likely just match on price and keep the account. That is also the return: the discount is paid for by the qualification work, not by a migration that may never happen. The sequencing follows from that. A team that funds a second serving path this quarter has something specific to point at when the contract comes up next quarter, while a team that waits until the renewal is asking for a concession on the strength of a model it has never served.
