What is cross-version KV cache reuse?

It means letting an updated model continue directly from the **KV cache** its previous version built for a shared prompt, instead of re-running prefill with the new weights. The arXiv paper 2611.14051 tests it across four same-family updates, with no mapping or selective recompute.

Shared system prompts and RAG documents are cached to save prefill time, and a model update usually means rebuilding those caches. Reuse kept median time-to-first-token speedups of **2.10–5.08×** on three longer-prefix QA benchmarks, but answer quality changed by benchmark and by update.

How does it relate to cross-model KV translation?

[Cross-model KV translation](/ai-explained/context-mobility-cross-model-kv-translation) trains a mapper so a different model can read a cache. Cross-version reuse skips the mapper and relies on the update keeping the same architecture and size, which is why it only applies between versions of one model.

Is matching the first output token enough to call reuse safe?

No. The paper found cases where reuse and native inference began with the same answer token but then differed in content or in where generation stopped. It measures **fidelity** separately with Functional Cache Divergence and **task utility** with accuracy or F1.

LLM· 2026-11-13

DeepSeek-V4-Flash KV reuse study — Cross-version KV cache reuse — What does it mean?

TL;DR

What is it: A November 2026 paper, Can KV Cache Be Reused Across Model Updates?, tests cross-version KV cache reuse: an updated model such as DeepSeek-V4-Flash-0731 continues directly from the KV cache its predecessor built for a shared prompt, with no mapping step and no recompute.
Why it’s needed: Shared system prompts and RAG documents are cached to skip prefill, and a model update usually means rebuilding every one of those caches; reusing them gave median time-to-first-token speedups of 2.10–5.08× on three longer-prefix QA benchmarks.
vs previous: Ordinary prefix caching reuses a cache only under the same weights, and cross-model KV translation trains a mapper to convert one; this paper reuses the old cache as is and shows that judging it needs two separate tests, task utility and fidelity, because one can pass while the other fails.

Jargon

KV cache: The keys and values a model computes for every token it has read, stored so later tokens do not recompute them. See the module.
Prefill: The pass that reads the whole prompt and builds its KV cache before the first output token. Skipping it is where cache reuse saves time.
TTFT: Time to first token: how long a user waits before any output appears. On long prompts, prefill is most of it.
Post-training update: A new checkpoint of the same model, with the same architecture and size, trained further (for example for agent or coding skills). The shapes stay the same; the weight values change.
Needle-in-a-haystack (NIAH): A retrieval test that hides many key-value records in a long prefix and asks the model for one value. In this paper each key and value is a single token.
Functional Cache Divergence (FCD): The paper's fidelity metric: the Jensen–Shannon divergence between the new model's next-token distributions on its own cache and on the old cache, scaled to 0 (identical) through 100.
Harmful flip rate: The share of answers the new model gets right on its own cache that become wrong when it uses the old cache.
Logit / answer margin: Logits are the model's raw token scores before softmax. The answer margin is how far the correct token's logit leads its strongest rival.

The news. On November 10, 2026, the arXiv technical report Can KV Cache Be Reused Across Model Updates? A DeepSeek-V4-Flash Case Study tested direct cache reuse across four same-family updates: GLM-5.2 → GLM-5.3, DeepSeek-V4.1-Flash → Flash-1017, DeepSeek-V4-Flash → Flash-0731, and Qwen3-30B-A3B → Instruct-2507. It compares three setups: the new model on its own cache (Target), the old model on its own cache (Source), and the new model continuing from the old model's cache (KV Reuse). It also introduces Functional Cache Divergence to measure how far reuse moves the new model's predictions. Read the paper →

Picture a restaurant that changes head chefs between lunch and dinner. The lunch chef left the kitchen full of prepped trays — onions diced, sauces reduced — for the dishes most people order. The dinner chef trained in the same kitchen but works from a revised recipe book. Re-prepping everything takes an hour; cooking straight from the trays takes no time at all. The trays were made to the old recipe book, so the real question is whether the new chef's dishes still come out right.

An updated model facing an old KV cache is in exactly that position. Every cached key and value was produced by the old weights, including the key and value projections, while the new model's queries and every layer that reads the cache now run on the new weights. A post-training update such as V4-Flash → Flash-0731 keeps the architecture and the size, so the cache has the right shape and the new model can run on it. Matching shapes make reuse possible; they say nothing about whether the answers survive.

Two inspectors could judge the dinner service. The first asks: did the diner get the right dish? That is task utility — the usual accuracy, or F1 (word overlap with the reference answer) for open-ended QA. The second asks: does it taste like what the new chef would have made from scratch? That is fidelity, and the paper measures it with Functional Cache Divergence (FCD): run the new model twice over the same following tokens, once on its own cache and once on the old one, and compare the full next-token probability distributions at each position. FCD catches a shift in probabilities even when both runs pick the same token.

The central finding is that the two inspectors disagree: a reused cache can keep answers correct while the model's predictions have measurably drifted. Drift also depends on the task. On the V4-Flash update, mean FCD on needle retrieval rose from 2.87 at a 512-token prefix to 30.62 at 32K, while ordinary text continuation on FineWeb-Edu stayed between 3.04 and 3.64 across the same lengths. Fetching one exact item from a long prefix is where the old cache hurts fidelity most.

The trays carry more than ingredients; they also carry the old chef's reading of the order. When the authors gave Flash's cache to the earlier Flash-Base checkpoint — which, on its own cache, usually restated the question and scored 0% on retrieval — its accuracy rose to 91.15% at a 2K prefix with its own weights unchanged. A KV cache stores how a model read the prompt, including how it decided to answer, not only the prompt's facts. For an update this cuts both ways. On natural-language QA, reuse changed the content of answers and the point where the model stopped, even in cases where both runs began with the same first answer token. Agreement on the first token is therefore not a fidelity test.

Across the four updates, the number that separates safe from unsafe reuse is the harmful flip rate on single-token retrieval. Read the V4-Flash row next to its accuracy column: reuse raised retrieval accuracy above Target at every measured length, yet about a quarter of the answers Target got right turned wrong at 32K, because other answers flipped the opposite way. An accuracy average can hide which answers changed, so a reuse check needs the paired, per-question comparison.

Update (source → target)	Harmful flip rate, 32K prefix	Retrieval accuracy under reuse	Source
GLM-5.2 → GLM-5.3	4.22%	Tracks Target through 16K; at 32K, 86.46% (Target) vs 90.62% (reuse)	arXiv 2611.14051
DeepSeek-V4.1-Flash → Flash-1017	0.53%	Near-native; Target and reuse both 97.92% at 32K	arXiv 2611.14051
DeepSeek-V4-Flash → Flash-0731	26.53%	Above Target at every length, but falls as the prefix grows	arXiv 2611.14051
Qwen3-30B-A3B → Instruct-2507	78.30%	Below Target at every length; largest gap at 32K	arXiv 2611.14051

Why do most answers survive the drift? Treat each answer as a race between the correct token and its closest rival, measured in logits. That lead is the answer margin. Hold three things fixed: the update (V4-Flash → Flash-0731), the task (needle retrieval at a 4K prefix, 192 queries), and the swap to the old cache. The paper measured a median margin change of −1.00 logit under reuse. Apply that shift to two answers (illustrative): one that led its rival by 3.0 logits drops to 2.0 and stays correct; one that led by only 0.6 drops to −0.4 and now loses. The real tally: 168 correct on the native cache (87.50%), 6 lost, 14 repaired, 176 correct under reuse (91.67%). Answers with a wide native lead absorb the drift; answers with a thin lead are the ones that flip, in either direction.

For a serving team, the payoff is time to first token. With the old cache already available, reuse cut median TTFT by 5.08× on HotpotQA, 2.10× on 2WikiMQA, and 2.51× on MultiFieldQA-en. Quality did not move with it: MultiFieldQA-en kept its F1, while HotpotQA and 2WikiMQA lost about 3 and 7 F1 points. On BABILong's shorter prefixes, loading the cache cost more than the prefill it replaced, for a 0.86× result — a slowdown — which echoes the short-prefix limits of ordinary prefix caching. The paper's results make cross-version reuse a per-update, per-workload call: speed, paired flips and fidelity each have to be measured, because none of them predicts the others.

Goes deeper in: LLM Serving → Prefix Caching & RadixAttention → When It Helps, When It Fails

Related explainers

Cross-model KV translation — the other way to carry a cache to a different model: train a translator instead of reusing it as is
Prefix-cache metastability — what happens to a prefix cache's hit ratio when a traffic burst fills it

Frequently Asked Questions

Check what you knowMap your AI & GPU knowledge across every track — free, role-based