# galaxy s26 ultra

Published articles for galaxy s26 ultra.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## HyCo: Hybrid Drafting for Cost-Efficient Mobile-Cloud Speculative Decoding

DevFeed: [HyCo: Hybrid Drafting for Cost-Efficient Mobile-Cloud Speculative Decoding](<https://devfeed.tech/articles/hyco-hybrid-drafting-for-cost-efficient-mobile-cloud-speculative-decoding-93736.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.21436>)

Author: Hanbin Bae, Heejin Choi, Joohyung Lee, Byoung Jin Choi, Sangjun Park, Hoon-Young Cho

Published: 2026-10-21T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Model Optimizer](<https://devfeed.tech/topics/model-optimizer.md>), [On-device AI](<https://devfeed.tech/topics/on-device-ai.md>), [quantization](<https://devfeed.tech/topics/quantization.md>)

Tags: [5g](<https://devfeed.tech/tags/5g.md>), [acceptance](<https://devfeed.tech/tags/acceptance.md>), [bench](<https://devfeed.tech/tags/bench.md>), [client](<https://devfeed.tech/tags/client.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [deep](<https://devfeed.tech/tags/deep.md>), [delivery](<https://devfeed.tech/tags/delivery.md>), [device](<https://devfeed.tech/tags/device.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [draft](<https://devfeed.tech/tags/draft.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [galaxy](<https://devfeed.tech/tags/galaxy.md>), [galaxy-s26-ultra](<https://devfeed.tech/tags/galaxy-s26-ultra.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [hidden](<https://devfeed.tech/tags/hidden.md>), [hybrid](<https://devfeed.tech/tags/hybrid.md>), [mobile](<https://devfeed.tech/tags/mobile.md>), [speculative-decoding](<https://devfeed.tech/tags/speculative-decoding.md>)

### AI overview

HyCo builds speculative drafts on a mobile NPU by using the server model's multi-token prediction head to form a deep token chain, then one small-language-model pass to add alternative branches. The server supplies the head's hidden state, input embedding, and key-value entries. Quantization, restricted vocabulary, and a 1K-token window fit both drafters to the device. An ε-threshold acceptance rule relaxes exact preservation of the verifier's sampling distribution. In single-client Galaxy S26 Ultra tests over LTE/5G, HyCo reduced server GPU time per output token versus server-side MTP self-speculation by 38.6% on Spec-Bench and 29.2% on 32K-token summarization. On Spec-Bench, it delivered 9.0 percentage points more saving and 1.9x the speed of an SLM-only drafter under the same acceptance rule. Its judge score differed from Cloud+MTP by +0.2 points, with a paired 95% CI from −0.9 to +1.3; exploratory translation favored SpecEdge-ε. Without the relaxed rule, HyCo reported savings of 27.7% and 11.9% on the two datasets.

### Source excerpt

arXiv:2610.21436v2 Announce Type: replace-cross Abstract: Device-cloud speculative decoding reduces server computation, but growing a deep draft tree with a small language model (SLM) requires repeated passes on a mobile neural processing unit (NPU). We present HyCo, which runs the server model's native multi-token prediction (MTP) head on the device to build a deep chain and uses one SLM pass to add alternative branches. The server returns the hidden state, input embedding, and key-value entries needed by the head instead of executing its sequential steps. Quantization, a restricted vocabulary, and a 1K-token window fit both drafters to the NPU. An {\epsilon}-threshold rule additionally relaxes exact preservation of the verifier's sampling distribution. In single-client runs with a Galaxy S26 Ultra over LTE/5G, HyCo reduces server GPU time per output token relative to server-side MTP self-speculation by 38.6% on Spec-Bench and 29.2% on 32K-token summarization. With the same acceptance rule, three runs of each system on Spec-Bench show 9.0 percentage points more saving and 1.9x delivery speed than an SLM-only drafter. HyCo differs from Cloud+MTP by +0.2 judge-score points (paired 95% CI: -0.9 to +1.3); exploratory translation favors SpecEdge-{\epsilon}. Without relaxation, HyCo exact saves 27.7% and 11.9% on the two datasets.