DeepSeek says its decoding trick makes generation 57–78% faster. The thread asks who's left to sell tokens to
DeepSeek published a paper claiming its DSpark speculative-decoding module makes per-user generation 57–78% faster on the same hardware. The replies split between admiration for the openness and a darker read on who can still IPO selling inference.
DeepSeek dropped a paper on DSpark, a speculative-decoding module bolted onto its V4 models, claiming 57–78% faster per-user generation at the same hardware capacity — and shipped the weights to Hugging Face the same day.
The thread's sharpest voice wasn't about the math. It was the warning that token-sellers eyeing an IPO have a market that's 'almost solely developers', open models a step behind, and no customer boasting higher profits from AI. Also noted: the headline isn't even the paper's title, and speculative decoding dates to 2022.
What DSpark actually claims
DeepSeek's paper, posted as a PDF on GitHub, describes a speculative-decoding module — a small fast model drafts tokens that the large model then verifies in bulk — wired into its DeepSeek-V4 line. The number doing the work, quoted in the thread from the abstract: "At matched system capacities, DSpark delivers 57% to 78% faster per-user generation." The weights went up immediately as DeepSeek-V4-Flash-DSpark and DeepSeek-V4-Pro-DSpark, and a user reported the Pro model already running at roughly a quarter of competitors' prices.
One detail a careful reader flagged: the submission's title is misleading. "It's the first line of the abstract instead of the paper title. Speculative decoding for LLM inference was published in 2022," pointing at the original arXiv work. DSpark is an improvement on a known technique, not a new genre — worth keeping straight before the hype compounds.
The "why now" everyone reached for
The recurring read in the thread was economic, not technical: this is how DeepSeek keeps prices low without taking on debt.
"This is just one of many papers DeepSeek have released to be able to serve models at extremely cheap prices, unlike the others taking on >$100B+ of debt in building data centers for the same thing." — a commenter with \~9.5k karma
A developer pointed at his own bill as evidence the cheapness is real: a month on DeepSeek-V4-Pro inside an agentic coding tool, 1.5 billion tokens, about $40. Another tied the paper to the price cut DeepSeek made weeks earlier — "presumably this has been in production for a while."
The part that should worry the labs trying to IPO
From there the thread turned to who's actually selling this. The bluntest case against the token-vendor business model:
"These companies providing tokens, whether SOTA or not, that want to IPO are so fucked as time goes on … a TAM that consists almost solely of developers, with no customer of theirs actually boasting increased profits as a result of AI. I fear their time to IPO may have passed." — the founder of a data consultancy
Underneath the price war runs a sharper grievance about who still shows their work:
"DeepSeek continues to not only push the boundaries but also publish these incredible papers explaining how they achieved their gains — something the American labs no longer do unfortunately." — a commenter with \~1.9k karma
That's the real reversal. A 57–78% speedup is a good engineering result; the thread treats it as a data point in a different argument — that the lab giving its methods away for free is squeezing the labs that need a trillion-dollar valuation to survive.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Meta's arbitrator fined a whistleblower $50,000 per criticism. She responded by sitting silent on stage for an hour
Sarah Wynn-Williams sat through an hour-long festival panel without saying a word, because Meta's private arbitrator had set a price of $50,000 for each criticism of the company. Meta then argued the silence itself was a violation.