While the AI world spent the weekend on OpenAI's gated, government-approved GPT-5.6, its biggest open-source rival made a quieter move in the opposite spirit. DeepSeek published a new technique for running its models faster — and then gave the whole recipe away. The headline doing the rounds, "60–85% faster," is real but easy to misread, so it's worth getting straight, because the way DeepSeek competes is the actual story.

What DeepSeek actually released

The drop, dated today, is a paper called DSpark plus an open-source repository, DeepSpec, containing the trained components and training code. It is not a new model — it's a faster way to serve an existing one, and it's built on a method called speculative decoding.

Speculative decoding is a clever, lossless speed trick. Normally a language model generates one token at a time, each requiring a full pass through the network — slow and expensive. Speculative decoding adds a small, fast "draft" model that guesses a whole chunk of upcoming tokens, which the big model then checks in a single pass, keeping the ones it agrees with. When the drafts are good, you get several tokens for the price of one — and because the big model verifies everything, the output is identical to what it would have produced anyway. DeepSeek's contribution is a smarter drafter design and a system that adjusts, on the fly, how aggressively to guess based on how busy the servers are.

Reading the "60–85%" honestly

Here's where the headline needs a footnote. DeepSeek reports that DSpark improves per-user generation speed by 60–85% for one model variant (and a more modest 57–78% for the larger one). But three caveats matter, and DeepSeek is admirably upfront about all of them:

• The figure is DeepSeek's own measurement, from its own production system. No outside party has verified it yet — the paper is hours old.
• It's measured against DeepSeek's own previous speed-up method (a scheme it calls MTP-1), not against a model with no optimisation and not against a competitor. It's "faster than our last trick," not "faster than everyone."
• It's a per-user gain "at matched throughput" — a measure of responsiveness under load, not a blanket "everything is now 85% faster." DeepSeek's own paper explicitly warns against reading the top number as a flat multiplier.

The honest summary: DeepSeek has meaningfully raised the speed ceiling of its own serving stack, validated it in live production — which most papers of this kind can't claim — and treat the exact percentage as a vendor figure until others test it.

Why this is the more interesting half of the weekend

Step back, and the contrast with the GPT-5.6 story is the point. This week showed two very different theories of how to win at AI. One — the US frontier labs', sharpened by Washington now approving who can even use the newest models — competes on raw capability and tightly controlled access. The other, DeepSeek's, competes on cost and openness: squeeze more output from the same hardware, drive the price-per-token down, and then publish the method so anyone can use it.

That second strategy is quietly relentless. DeepSeek's models already undercut Western rivals dramatically on price; a serving technique that lifts per-user speed by 60–85% (on DeepSeek's own numbers) without new chips pushes that cost advantage further — and giving it away pressures everyone else's margins, not just its own. It's the same competitive logic running underneath the whole open-vs-closed divide we've tracked: where the frontier labs guard capability, the open camp attacks economics, in public.

The takeaway

Don't read this as "DeepSeek's model got 85% better." Read it as a snapshot of strategy. The same weekend the US government turned a frontier model into a controlled, approval-only release, a Chinese lab took its own model, made it cheaper to run, and handed the instructions to the world. Whether "open and efficient" or "powerful and gated" wins is the defining contest in AI right now — and this weekend, both sides showed their hand at once.