Research

Speculative Decoding Adoption

How speculative decoding works, what speedups it delivers in practice, why acceptance rate determines everything, and how widely it is deployed across serving stacks in 2026.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: July 2026

Speculative decoding is the rare optimisation that makes inference faster without changing the output distribution. That property, exact equivalence rather than approximation, is why it moved from research to default in serving stacks faster than most efficiency techniques.

The Mechanism

A small fast draft model proposes several tokens ahead. The large target model then verifies all of them in a single forward pass, because verifying a sequence is parallel work while generating one is sequential. Accepted tokens are kept; at the first rejection, generation falls back to the target model's own choice and drafting restarts.

The output is mathematically identical to what the target model would have produced alone. You are not trading quality for speed, which is unusual and is the whole appeal.

Acceptance Rate Decides Everything

Acceptance rateTypical outcome
High, easy or predictable textSubstantial speedup, often 2-3x
ModerateModest speedup
Low, hard or unusual contentNet slowdown from wasted draft compute

Speedup depends on how often the draft model guesses right, which depends on content. Boilerplate code, formulaic prose, and structured output draft extremely well because the next token is often nearly determined. Novel reasoning and unusual domain content draft poorly. Reported average speedups therefore hide wide variance, and a figure measured on code completion will not hold on open-ended reasoning.

The draft model also has to be genuinely cheap. If drafting costs a meaningful fraction of verification, the arithmetic stops working even at good acceptance rates.

Variants

Three families. Separate draft models use a smaller model from the same family, which is simple but requires a suitable small sibling. Self-speculation methods such as Medusa and EAGLE add lightweight prediction heads to the target model itself, removing the need for a separate model. And n-gram or prompt-lookup drafting predicts from text already in the context, which is remarkably effective for tasks with heavy repetition such as editing a document or summarising a provided passage.

Prompt-lookup drafting is underrated. For any workload where the output substantially echoes the input, it is nearly free.

Adoption

Speculative decoding is supported across the major open serving stacks and is used internally by large providers, though most do not disclose where. Since it does not change output, providers have no obligation to announce it, which makes deployment hard to track from outside. Treat it as widely deployed rather than precisely measurable. See the serving stack comparison.

Brand Visibility Implications

The indirect effect is what matters. Cheaper inference means systems can afford more steps within the same budget: more retrieval, more comparison, more verification. Techniques like this are part of why agentic, multi-step answering became economically viable, and multi-step answering cites more sources than single-pass answering does. See the inference and training split.

Methodology

Compiled from provider documentation, published benchmark suites, and independent measurement through July 2026. Inference performance figures depend heavily on model, quantisation, batch size, context length, and region, so ranges are given rather than point values and cross-provider comparison carries real uncertainty. Re-verify before quoting. Updated quarterly.

How Presenc AI Helps

Presenc AI measures how many sources reach an answer across platforms, which is where inference-efficiency gains ultimately show up for brands.

Frequently Asked Questions

A technique where a small fast draft model proposes several tokens ahead and the large target model verifies them in a single forward pass, because verification is parallel while generation is sequential. The output is mathematically identical to what the target model would have produced alone.
No. It produces output identical in distribution to the target model generating alone, because rejected draft tokens are replaced by the target model's own choice. That exact-equivalence property is why it was adopted faster than approximate efficiency techniques.
It depends entirely on acceptance rate, which depends on content. Predictable text such as boilerplate code and structured output can see 2-3x. Novel reasoning and unusual domain content draft poorly and can produce a net slowdown from wasted draft compute. Average figures hide wide variance.
A variant that predicts upcoming tokens from text already present in the context rather than from a draft model. It is remarkably effective and nearly free for workloads where output echoes input, such as document editing or summarising a provided passage, and it is underused relative to how well it works.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.